How Visible Are Indian Brands in AI Search?

A benchmark of how often, and how well, Indian brands are cited when AI assistants answer category questions, and what separates the sourced from the merely asserted

Melivana | PR Intelligence Series 2026, Report 6 of 8

Executive summary

Search is no longer a page of ten blue links. Increasingly, it is a paragraph. When an Indian consumer asks "which is the best UPI app for small businesses," or a procurement head types "top HR software for mid-market companies in India," or a founder asks "who are the leading contract-manufacturing partners for consumer electronics," the answer that shapes the decision is now, more often than not, generated. It is written by ChatGPT, by Google's Gemini and AI Overviews, or by Perplexity, synthesised in seconds from a small set of sources the model has decided to trust, or from no live sources at all.

This is the central shift AEO, Answer Engine Optimization, is built to address. AEO is the discipline of getting a brand named, and ideally cited with a source, inside the answer an AI assistant returns. Its sibling, GEO (Generative Engine Optimization), covers the broader craft of engineering content and reputation so generative systems surface you. In 2026 these are no longer speculative disciplines. Google reported that AI Overviews crossed two billion monthly users across more than 200 countries, and that AI Mode, the fully conversational search experience, passed 100 million monthly active users, with the United States and India named as its two anchor markets. Gemini's consumer app crossed 750 million monthly users early in the year and kept climbing. In India specifically, Gemini leads ChatGPT on app market share and the country carries one of the highest chatbot download shares in the world. The answer layer has arrived in India at scale, and it arrived fast.

The question this report answers is deceptively simple: when an AI assistant answers a commercial category question relevant to India, does an Indian brand actually show up, and show up cited, with a source the user can verify, rather than merely asserted from memory?

To answer it we did not model. We ran a real audit. On 8 July 2026, Melivana's PR Intelligence unit posed 20 India-context "best/top X in India" questions across five sectors, Fintech, SaaS, D2C, Healthcare, and Manufacturing, to three live assistants: ChatGPT (gpt-4o), Google Gemini (2.5-flash), and Perplexity (sonar-pro), each with web search enabled. That produced 60 answers, 682 brand mentions, and 334 citation instances spanning 237 distinct domains. Every figure in this report is a genuine count from that corpus, not a calibrated estimate.

Our headline finding is this:

Across the five sectors, 66.7% of category AI answers name an Indian brand together with a supporting source, and 79.5% of all brand mentions occurred in answers that carried a source. Meanwhile, every single answer named at least one Indian brand: presence is total. The real contest is not whether Indian brands appear, but whether the answer that names them can be traced.

That framing inverts the usual AEO anxiety. In India-context queries, Indian-brand presence turned out to be near-total, 0% of answers named no Indian brand. The battle is not for a seat at the table; the domestic field is already at the table. The battles that decide outcomes are sourcing (is the mention grounded and verifiable?) and concentration (do a few names monopolise the answer, or is it a long, flat list?). And the single most striking behaviour in the whole audit sat inside the sourcing question, which we return to repeatedly below: one major assistant answered every question closed-book, naming roughly nine Indian brands per answer with zero citations, pure, unsourced assertion at scale.

This report explains the 66.7% number, decomposes it into five findings, breaks it down by sector, situates it against the real citation landscape our sibling reports dissect, and closes with a practical AEO playbook built for the conditions Indian brands actually face, not thin presence, but thin sourcing.

A note on method and scope. This is Report 6 of an eight-part PR Intelligence Series, and it anchors a three-report AI trilogy. Reports 6, 7, and 8 all analyse the same round of AI querying, viewed through three different lenses. This report (6) measures brand visibility and citation quality, how often Indian brands are named with a source. Report 7, the India Media Citation Index, examines which publications and domains the assistants leaned on when they did cite. Report 8, the AI Answer Share report, converts the same corpus into competitive share-of-answer league tables. The three share one query set, the same five sectors, and the same date, so figures reconcile across the trilogy by design. The audit is deliberately bounded, we say exactly how bounded in the Limitations section, but it is real, repeatable, and meant to be run on a cadence rather than once.

Why AI search matters now for Indian brands

For most of the last two decades, the Indian visibility contest was a search-engine-results-page contest. Rank on page one of Google for a high-intent query, and traffic, and with it, consideration, followed. That world is dissolving at the top of the funnel. When an AI Overview or a chatbot answers the question directly, the user frequently never scrolls to the blue links at all. Industry measurement in 2026 put AI Overviews' appearance rate as high as roughly 60% of US queries and materially lower but rising globally. The mechanical consequence is a compression of the discovery surface: where a search engine returned ten organic results, an AI answer typically leans on a far smaller set of cited domains, when it cites at all. The shelf just got shorter, and only the brands the model trusts get to stand on it.

Three forces make this acute for India specifically.

First, adoption is disproportionately high and disproportionately young. India is one of the largest markets by users for both Gemini and ChatGPT, and its user base skews toward exactly the digitally native, mobile-first cohort that companies most want to reach. When a 26-year-old in Pune or a small-business owner in Surat asks an assistant to shortlist a lending app or an accounting tool, that answer is the top of the funnel.

Second, and this is where our audit reframes the received wisdom, the problem is not that Indian brands are absent. A common assumption in the AEO literature is that Indian challenger brands are structurally under-represented in AI answers because their Wikipedia, press, and review footprints are thinner than Western peers'. On the narrow question of presence in India-context queries, our data flatly contradicts that fear: every one of the 60 answers named at least one Indian brand, and the answers were, if anything, over-full, long lists of eight to fourteen domestic names apiece. The structural thinness of the Indian corpus does not erase Indian brands from the answer. It shows up somewhere subtler: in whether the answer that names them is sourced, and in which sources the models reach for when they do cite.

Third, the categories where India is building fastest are the categories AI treats most conservatively. Fintech and healthcare fall under what AI platforms treat as "your money or your life" (YMYL) territory, where trust filters are strictest. Our audit shows this not as absence but as concentration and sourcing pressure, healthcare, as we will see, was the most concentrated sector in the whole study, and the least-sourced by mention share.

The upside is that AI visibility is winnable, and it is winnable through disciplines Indian communications teams already own, earned media, reputation, structured content, more than through paid spend. The rest of this report quantifies the current state from real data and maps the route.

How LLMs choose what to name and what to cite

To interpret the benchmark, it helps to be precise about the machinery, because our audit surfaced two genuinely different behaviours from the three engines, and the difference is the story.

An AI assistant's decision to name a brand and its decision to cite a source for that name are separate acts, governed by four overlapping mechanisms.

1. Parametric (training) knowledge. Models carry a compressed memory of their training corpus. Brands that appeared frequently, consistently, and in reputable contexts are "known" to the model and surface even without live retrieval. This is why brand salience is one of the strongest single predictors of LLM visibility in published work. Familiarity, accumulated over years of coverage, is baked in, and, crucially, it is enough on its own for a model to name a brand without ever consulting a live source. Our closed-book engine is the pure expression of this channel.

2. Retrieval (RAG). Assistants increasingly ground answers in live search. But "increasingly" is not "uniformly," and our audit caught the divergence in the act. Gemini (2.5-flash) and Perplexity (sonar-pro) browsed and cited sources on 100% of their answers. ChatGPT (gpt-4o), with web search nominally enabled, nonetheless answered closed-book on 100% of its answers, zero citations, every time. Retrieval is the fastest-moving citation lever, but it is only pulled if the engine actually retrieves. When it does not, the brand names still appear; they simply arrive ungrounded.

3. Source authority and consensus. When models do retrieve, they privilege sources they treat as trustworthy and, decisively, they privilege consensus. When multiple independent, reputable sources say the same thing about a brand, the model treats it as established and repeats it. This is why earned and third-party media dominate AI citations. But which sources count as "trustworthy" for commercial queries turned out to be surprising in our corpus, and we detail it below and in Report 7: it was not the business press.

4. Structure and freshness. Given a shortlist of candidate sources, models preferentially lift from content that is machine-extractable and current, a direct answer under a question-shaped heading, explicit statistics, clear provenance, structured data. This lever governs which page within a domain gets cited once the domain is in the retrieval set.

These four mechanisms interact, but the audit teaches one lesson above all: naming and sourcing are decoupled. A brand can be named on the strength of channel 1 alone, with channels 2 to 4 never engaged. The commercial value of a mention lives almost entirely in whether the sourcing channels fired. That is why our central metric is not "was a brand named", in India-context queries the answer is essentially always yes, but "was it named with a source."

The headline: 66.7% of category AI answers name an Indian brand with a source

Our central metric is deliberately strict, and the audit lets us apply it against real counts rather than estimates. We do not count an answer as a sourced "hit" merely because it mentions an Indian brand, because by that loose standard the audit scores 100%, and 100% flatters away the real problem. We count an answer as sourced only when the assistant names an Indian brand and the answer carries an identifiable source, an inline citation or a linked domain, behind it.

By that standard, 40 of the 60 answers, or 66.7%, name at least one Indian brand with a source. The other 33.3% name Indian brands with no verifiable source at all.

The 66.7% is not a coincidence of content quality; it is almost entirely a product of engine behaviour. Two of the three assistants cited on every answer; one cited on none. Two engines × 20 questions = 40 sourced answers; one engine × 20 questions = 20 unsourced answers. The headline rate is, in effect, the market-share-weighted average of "does this engine browse." That is a profound finding in its own right: in this corpus, whether a brand mention is verifiable depended less on the brand and more on which assistant the user happened to ask.

A second, complementary metric sharpens the picture. Because answers vary in how many brands they name, we also measured what share of all 682 brand mentions landed in sourced answers. That figure is 79.5%, higher than the 66.7% answer-level rate, because the browsing engines tended to produce the longer, denser lists. Four in five brand mentions in the corpus sat inside an answer a user could trace back to a source. That is genuinely encouraging. But the inverse is the risk: one in five brand mentions, and one in three answers, carried no source at all.

The number to hold in mind, then, is not "66.7% good" or "33.3% bad." It is this: in India-context category queries, Indian brands are essentially always present, but a third of the answers that name them cannot be verified, and that unverifiable third is concentrated in a single engine's behaviour. That is the gap this report exists to explain, and the five findings below dissect it.

Key finding 1, Two-thirds of answers name an Indian brand with a source; presence itself is total

The headline is also the first finding, because it reframes how Indian marketers should think about AI visibility. The instinct, imported from the SEO era and from Western AEO commentary, is to ask "am I even in the answer?" Our audit says: for India-context queries, your category almost certainly is. Every one of the 60 answers named a domestic brand; 0% of answers named no Indian brand. The presence war, in this query type, is already won.

That means the strategically useful question is not presence but sourcing: is the answer that names your category grounded? Here the real number is 66.7% of answers sourced, with 79.5% of mentions landing in sourced answers. The one-third of answers that are unsourced are not "missing Indian brands", they are full of Indian brands, asserted without a trace. That is a different problem than the literature usually describes, and it demands a different response: not a fight to be included, but a fight to make the inclusion verifiable and correct.

There is also a quality gradient inside the sourced two-thirds. A mention backed by an independent, durable source survives across engines and re-runs; a mention resting on a single lightweight page appears once and vanishes on the next run. Our sibling Report 7 shows that for these commercial queries the cited sources skewed heavily toward community and social platforms rather than sturdy business press, which makes durability a live concern even within the sourced band. This is precisely why the audit is designed to be re-run on a cadence rather than treated as a one-time snapshot: a mention that is sourced today on Gemini and gone tomorrow is not a stable asset.

Implication. For most Indian brands in most of these categories, the first job is not to appear, you already do, but to ensure the answers that name you are sourced, and sourced to something durable. Presence is table stakes; verifiable presence is the actual game.

Key finding 2, Concentration is low, because answers are long lists, the top three capture only 5.5% to 20%

If the first finding is about sourcing, the second is about the shape of the answer, and it overturns another piece of received wisdom. AEO commentary often warns that AI answers are winner-take-most: that two or three names monopolise every response. In our India-context audit, the opposite was true. The answers were long, flat lists, and the top three brands in a category captured only a small slice of all Indian-brand mentions, ranging from 5.5% in D2C to 20.0% in Healthcare.

The reason is mechanical. Because these were "best/top X in India" discovery queries, the browsing engines in particular returned expansive rosters, the average answer named between 8.5 (Fintech) and 13.75 (D2C) brands. When an answer lists a dozen names, no single name can command a large share, and even the top three together capture a minority. The category leaders bear this out: the most-cited brand in each sector held only a modest share of all mentions, PhonePe 2.9% in Fintech, Zoho CRM 2.2% in SaaS, Mamaearth 1.8% in D2C, Tata Steel 3.7% in Manufacturing. The single exception, and it is instructive, is 1mg in Healthcare at 7.8%, in a sector where the lists were shorter and the same few names recurred.

This flips the strategic reading. Low concentration is, on one hand, opportunity: the answer has room for many names, so the marginal cost of becoming one more name on a fourteen-brand list is low, and being "in the answer" does not require unseating an entrenched leader. On the other hand, low concentration is dilution: being one of fourteen names carries far less decisive weight than being one of three. In a long, flat list, mere presence is cheap and therefore worth less; what matters is ordinal position and framing, whether you are named first, described as the category leader, or buried mid-list.

Healthcare is the exception that proves the rule, and it points to where the game is different. Its top-3 share of 20.0%, nearly four times D2C's, reflects shorter lists and stronger consensus around a small set of names (1mg, PharmEasy, Apollo Pharmacy). In YMYL categories, models are more conservative, name fewer options, and lean on the same trusted incumbents. There, concentration is real and top-three position genuinely matters. In the consumer and B2B list-categories, the fight is instead to rise within a crowded, low-concentration roster.

Implication. Do not import the "winner-take-most" assumption into India-context discovery queries, the data does not support it. Treat these long-list categories as low-concentration environments where presence is easy but prominence is scarce, and treat Healthcare (and YMYL generally) as the genuinely concentrated exception where top-three consensus is the prize. Report 8 converts exactly this concentration structure into competitive share-of-answer league tables.

Key finding 3, Coverage drives citation: the sourced engines cited, the closed-book engine did not

The third finding names the lever, and the audit hands us an unusually clean natural experiment to demonstrate it. When we ask why two-thirds of answers were sourced and one-third were not, the answer is not brand quality or category maturity, it is whether the engine went and looked. Gemini and Perplexity browsed the live web and cited sources on 100% of answers. gpt-4o did not browse and cited on 0%. The presence or absence of a source tracked, perfectly, with the presence or absence of retrieval.

This is the coverage-to-citation link rendered in its starkest form. Retrieval is the act of a model consulting the coverage that exists about a brand, the third-party pages, community threads, and reference entries that make up its earned footprint. When the engine retrieves, that coverage becomes the source it cites; when it does not, the same brand is named from parametric memory with nothing attached. The lesson generalises: a brand's earned coverage can only be cited if (a) it exists and (b) the engine retrieves it. Indian brands can influence (a) directly and (b) indirectly, by making their coverage the kind that retrieval-first engines surface.

The mechanism underneath is consensus. A brand mentioned once, in one place, is a data point a model can discount. A brand mentioned repeatedly, across independent, reputable sources, becomes, to a browsing model, an established fact about the category, and established facts get repeated and cited. This is why the browsing engines converged on recognisable domestic leaders in each sector while still listing a long tail: the leaders had the densest, most consensus-forming coverage, so they surfaced first and most confidently.

Two operating principles follow.

Consistency beats intensity. Retrieval engines weight freshness and repetition, so a steady cadence of coverage outperforms a single burst of equivalent volume. For Indian communications teams, this reframes PR from campaign-shaped spikes toward an always-on earned-media cadence that keeps the brand continuously present in the sources a browsing assistant retrieves on any given day.

Multi-source beats single-source depth. Ten mentions in one friendly outlet build far less citation equity than the same brand appearing across ten independent outlets, because consensus requires independence. Diversifying the earned footprint is what converts coverage into durable citation.

Implication. The coverage-to-citation link is the empirical heart of AEO, and our audit shows it in the cleanest possible form: sources appear exactly where retrieval happens, and retrieval cites exactly the coverage that exists. AEO is therefore not primarily a technical exercise; it is a PR and reputation exercise measured on a new surface. Build the consistent, multi-source coverage that browsing engines can find and consensus can reward, because on the engines that browse, that coverage is the citation.

Key finding 4, Answer-first, structured content is what actually gets lifted once an engine browses

The fourth finding shifts from whether coverage is retrieved to which page within the retrieved set gets cited. Once Gemini or Perplexity decided to browse a category, they did not cite every relevant page, they lifted from the ones that were machine-extractable, answer-shaped, and current. Holding a brand's earned footprint roughly constant, the format of the content it and its advocates publish materially changes how often an engine can extract and attribute it.

This matters because our audit's citation instances, 334 of them across 237 domains, were not distributed evenly across all available content about these brands. They clustered on pages that made the model's job easy: pages that answered the question directly, listed named options with clear provenance, and were freshly maintained. The published GEO literature is emphatic and consistent on the pattern, and our corpus is consistent with it: AI preferentially lifts a clear, self-contained answer placed right under a question-shaped heading; a large share of citations are drawn from the first third of a page; and the most-cited pages are overwhelmingly recently updated.

"Answer-first, structured content" is a concrete, teachable specification, not a vibe. It means:

• A direct answer immediately after a question-shaped heading. Lead with the answer, the named brands, the ranked list, the verdict, then elaborate. Browsing engines lift the lead.
• Extractable provenance. Specific statistics, named sources, dates, and short quotable sentences give a model machine-verifiable material to lift and attribute. Adjective-heavy marketing copy gives it nothing to cite.
• Structured data. FAQPage, Article/BlogPosting, Organization, and Product schema make a page's meaning and freshness machine-legible and are repeatedly associated with higher citation appearance rates.
• Freshness signals. The most-cited pages are recently updated; maintaining and visibly dating cornerstone content keeps it eligible for retrieval.
• Self-contained, chunked passages. RAG systems retrieve passages, not whole documents. Content written in modular, standalone chunks, each of which makes sense lifted out of context, is far more citable than reference-dependent prose.

For Indian brands the structural lever is doubly valuable, because it is the fastest-acting one available. Earned reputation compounds over quarters; parametric familiarity over training cycles. Content structure can be fixed in weeks, and it improves the citability of both owned properties and, via digital PR, contributed articles, and structured data on partner and community sites, the earned sources that carry the brand. Given that our sourced engines lifted so heavily from community and social platforms (Finding 5 and Report 7), the highest-leverage application of answer-first structure is often off the brand's own domain: making the third-party pages that name you as clean, current, and quotable as possible.

Implication. Structure does not decide whether an engine browses, that is the engine's behaviour (Finding 3). But once it browses, structure decides what it lifts. Audit your cornerstone content and the contributed and community content you can influence against the answer-first specification, because in a low-concentration long-list environment (Finding 2), being the cleanest, most quotable entry is often what elevates you from mid-list mention to named leader.

Key finding 5, In a third of answers, Indian brands are named with NO source, an entire engine, running closed-book

The fifth finding is the risk the whole audit was, in retrospect, built to expose, and it is not a rounding-error edge case, it is a third of the corpus and an entire engine. ChatGPT (gpt-4o) answered all 20 questions closed-book: it named Indian brands, roughly nine per answer, with zero citations, every time. That single engine's behaviour is the whole of the unsourced 33.3% of answers. It named domestic brands confidently, in ranked lists, across every sector, and attached nothing a user could check.

This is double-edged, and both edges cut.

The opportunity edge. An unsourced mention proves the model already associates the brand with the category, the parametric familiarity is present without any retrieval needed. That is latent equity. It means the brand has accumulated enough historical coverage to be "known." Convert it by supplying the consensus a browsing engine would need: earn the third-party coverage and publish the structured, citable content so that the retrieval-first engines also name you, with a source attached. Closed-book mentions are proof of brand memory; the goal is to make that memory verifiable everywhere else.

The risk edge, and it is serious. A closed-book mention is ungoverned. It can be wrong, an outdated product claim, a misattributed feature, a hallucinated statistic, a stale price, a competitor's attribute pinned to your name, and because it carries no source, the user cannot check it and the brand cannot easily trace it. Nine unsourced brand names per answer, produced by a major assistant across every query we posed, is a large surface of unverifiable assertion. In YMYL categories, Fintech, Healthcare, an unsourced, inaccurate claim about a financial product or a health service is not merely a marketing problem; it is a compliance and reputational exposure. And because closed-book output is generated from parametric memory rather than live sources, it is exactly the output least likely to reflect a brand's current, correct positioning.

The strategic response to the closed-book band is therefore twofold: monitor it continuously, which requires exactly the multi-engine audit design this report is built on, because you cannot see what a closed-book model is saying about you without deliberately querying it, and feed the parametric channel over time, by building the durable, consistent coverage that shapes what the next model training cycle "knows." You cannot cite-check a closed-book answer in the moment; you can only make the underlying memory it draws on more accurate through sustained, consistent presence in the corpus.

Implication. AEO is not only about increasing citations; it is about governing representation, including representation that arrives with no citation at all. A major assistant naming nine Indian brands per answer with zero sources is the clearest demonstration in this audit that "being named" and "being safely represented" are different things. Closing the gap between the closed-book assertion and the verifiable, correct claim is both the clearest growth path and the clearest risk-management priority.

Sector benchmark table

The table below summarises the real audit across the five sectors that anchor this entire trilogy. All figures are drawn from the 60-answer corpus of 8 July 2026 (12 answers per sector: 4 questions × 3 assistants). "Answers citing an Indian brand with source" applies the strict standard (named with an identifiable source). "Top-3 share" is the proportion of in-category Indian-brand mentions captured by the three most-mentioned brands. "Answers with no Indian brand" is the share of category answers naming no domestic brand at all.

Read the table as a map of a single, uniform structure with one meaningful axis of variation. The 66.7% sourced rate is identical across all five sectors, because it is driven by engine behaviour (two of three engines browsed) rather than by anything sector-specific. The 0% no-brand rate is also uniform: in India-context queries, every sector's domestic field is fully present. The one column that genuinely varies is top-3 share, and it tells the concentration story: D2C (5.5%) and SaaS (6.6%) are the flattest, longest-list categories where presence is cheap and prominence scarce; Manufacturing (10.4%) and Fintech (8.8%) sit in the middle; and Healthcare (20.0%) stands apart as the genuinely concentrated category, where a small set of trusted names, 1mg, PharmEasy, Apollo Pharmacy, command a fifth of all mentions and top-three position is worth fighting for.

Sector deep dives

Fintech. India's fintech story, UPI ubiquity, a dense neobanking and lending field, a globally-watched payments stack, gives the category rich earned media, and the browsing engines drew on it to produce moderately long lists (8.5 brands per answer on average, the shortest of the consumer-facing sectors). PhonePe led at 2.9% of mentions, with the top three, PhonePe, Google Pay, Paytm, capturing 8.8%. Sourcing followed the corpus norm: 66.7% of answers sourced, 76.5% of mentions in sourced answers, the lowest mention-sourcing rate among the digital sectors, a reminder that even a media-rich category is only as verifiable as the engine querying it. Fintech is also YMYL, so governance matters most here: an unsourced, inaccurate claim about a financial product carries regulatory weight, and a third of Fintech answers came with no source at all.

SaaS. Indian SaaS produced the second-longest lists (11.33 brands per answer) and the second-flattest concentration, Zoho CRM led at just 2.2%, and the top three (Zoho CRM, Freshsales, LeadSquared) captured only 6.6%. It also had the highest mention-sourcing rate of any sector at 83.8%, meaning that when SaaS brands were named, they were unusually likely to sit in a sourced answer. The picture is a crowded, well-sourced, low-concentration roster: for India-context SaaS queries, being in the answer is easy and usually verifiable, but no single name dominates, so the fight is for prominence and framing within a long domestic list rather than for inclusion.

D2C. Consumer D2C produced the longest lists of all, 13.75 brands per answer, and the lowest concentration, a top-3 share of just 5.5% (Mamaearth 1.8%, then Plum and Minimalist). Mention-sourcing was strong at 80.6%. D2C lives on exactly the surfaces the browsing engines retrieved, lifestyle and product media, review aggregators, marketplace listings, community discussion, so answers were dense and mostly grounded. But the flatness is the challenge: in a fourteen-name list where the leader holds under 2%, mere presence is nearly costless and nearly weightless. Durability and prominence, anchoring on sturdier reviews and earned coverage, and earning the top-of-list framing, are what separate a brand that is merely listed from one that is recommended.

Healthcare. Digital health, diagnostics, and health services behaved differently from every other sector, and the difference is the finding. Lists were shorter (9.58 brands per answer) and concentration was by far the highest, a top-3 share of 20.0%, with 1mg alone at 7.8%, followed by PharmEasy and Apollo Pharmacy. This is the YMYL conservatism showing up as consensus: models name fewer options and lean on the same trusted incumbents. Healthcare also had the lowest mention-sourcing rate at 74.8%, the most concentrated and the least-sourced sector, a combination that raises the stakes on both prominence and governance. For Indian health brands, top-three consensus is both harder to enter and more valuable once held, and the unsourced-claim risk is at its most acute here.

Manufacturing. Industrial and manufacturing, often assumed to be the visibility desert, was in fact one of the most populous, 13.67 brands per answer, second only to D2C, with Tata Steel leading at 3.7% and the top three (Tata Steel, L&T, Bharat Forge) at 10.4%. Mention-sourcing was solid at 79.9%. The assumption that B2B industrial brands are absent from AI answers did not hold for India-context queries: the browsing engines produced long, sourced rosters of domestic manufacturers. The concentration sat in the middle of the pack, a handful of storied names (Tata, L&T, Bharat Forge) anchored the top, but a deep tail of specialists filled out the list. The opportunity here is prominence within a surprisingly crowded field, and the durability of the sources, heavily aggregator- and directory-driven, per Report 7, is the variable to watch.

The real citation landscape: what the sourced engines actually cited

A benchmark of whether answers are sourced is incomplete without asking what they were sourced to, and here our audit produced its most counterintuitive result, one that Report 7 dissects domain by domain across all 334 citation instances and 237 distinct domains. For these commercial, "best X in India" queries, the sources the browsing engines cited were not traditional business press. They were, overwhelmingly, community, social, reference, and commercial-aggregator surfaces: YouTube, Reddit, LinkedIn, Instagram, Wikipedia, brand-owned domains, and SEO aggregators and market-research sites.

That is a structurally important fact for AEO strategy. It means the sources that decide whether an Indian brand is cited, and how it is described, for a commercial query are largely the sources a brand's PR team has historically not prioritised. The business and trade press still matter for parametric familiarity and long-term reputation, but in the live-retrieval moment for a "top D2C brands in India" or "best online pharmacy in India" query, the model reached for a Reddit thread, a YouTube roundup, a LinkedIn post, a Wikipedia entry, or a listicle on an SEO aggregator far more readily than for a business daily.

The AEO implication is direct and slightly uncomfortable. To be cited, not merely named, on the browsing engines for commercial queries, Indian brands must earn genuine, substantive presence on exactly these community and aggregator surfaces: credible Wikipedia entries where notability supports them; authentic, positive discussion on Reddit and LinkedIn; a footprint in YouTube reviews and roundups; accurate, current listings on the market-research and aggregator sites that models treat as convenient consensus. This is not an invitation to fabricate, consensus built on thin or manipulated sources is fragile and reputationally dangerous, but a redirection of earned-media effort toward the surfaces that actually feed the answer. Report 7 quantifies this landscape in full; the takeaway for this report is that sourcing strategy must follow the sources the engines actually use, and for commercial India queries those are social and aggregator platforms, not the newsroom.

The earned-media / share-of-voice link to AI citations

The through-line connecting all five findings is that AI citation is downstream of the coverage that exists and gets retrieved. Our natural experiment made this vivid: the engines that browsed cited the coverage they found; the engine that did not browse cited nothing while still naming brands from memory. Where retrieval fired, earned and third-party coverage became the citation. Where it did not, only parametric memory spoke.

This makes share of voice in the retrievable, earned corpus a leading indicator of share of voice in AI answers, and it is why this trilogy is designed as it is. Report 7 measures which sources the assistants actually leaned on, and finds, as above, a social-and-aggregator-dominated landscape rather than a press-dominated one. Report 8 measures the resulting share of AI answers, converting the low-concentration long lists of Finding 2 into competitive league tables. Report 6, this report, sits at the hinge, measuring how effectively coverage converts into sourced citation and how much of the answer arrives unsourced. The strategic reading is direct: a brand cannot durably buy its way into AI answers, but it can earn its way in, by building consistent, credible, multi-source coverage on the surfaces the browsing engines retrieve, and by shaping, over time, the parametric memory that the closed-book engines assert from. For Indian brands, whose presence in these queries is already total, the opening is not to appear but to make that appearance verifiable, prominent, and correct.

The closed-book risk, revisited

It is worth dwelling on why the closed-book condition (Finding 5) deserves board-level attention and not just marketing-team attention. An assistant that names nine Indian brands per answer with no source is making a large volume of assertions the brand did not author, cannot see without deliberate monitoring, and cannot correct in the moment because there is no source to counter. At best it is warm parametric equity, proof the brand is remembered. At worst it is an ungoverned narrative carrying factual errors into high-intent conversations at scale. In regulated YMYL categories, Fintech and Healthcare, where a third of answers were likewise unsourced, the exposure is acute.

The correct posture is neither to celebrate closed-book mentions as free visibility nor to ignore them, but to treat them as a monitored surface to be shaped over time. Because they cannot be cite-checked in the moment, the levers are indirect: query the closed-book engines deliberately and regularly to see what they assert; feed the broader corpus with consistent, accurate, structured coverage so the memory the model draws on is correct; and enforce message and fact consistency everywhere the brand appears, so that even an ungrounded assertion is more likely to be right. This is only possible with the repeated, multi-engine audit cadence that turns AEO from a one-time optimisation into an ongoing intelligence function.

The AEO playbook for Indian brands

Synthesising the real findings into an operating plan, here is the sequence we recommend for Indian brands, roughly in priority order. Note how it differs from a generic AEO playbook: because presence is already total in India-context queries, the emphasis shifts from getting into the answer to making the answer verifiable, prominent, and governed.

1. Measure sourcing, not just presence (baseline first). Stand up a repeatable audit, the design in this report, across ChatGPT, Gemini, and Perplexity (and Google AI Overviews), run on a regular cadence. Track the metrics that actually vary: sourced-citation rate, mention-sourcing share, top-3 concentration, and closed-book (unsourced) mentions for your brand and category. Presence will read near-100% and tell you little; sourcing, concentration, and the closed-book band are where the signal is.

2. Earn presence on the sources the engines actually cite. Per the real citation landscape and Report 7, commercial India queries are sourced to YouTube, Reddit, LinkedIn, Instagram, Wikipedia, brand domains, and SEO/market-research aggregators, not the business press. Redirect earned-media effort toward genuine, substantive presence on those surfaces, because those are what the browsing engines retrieve and cite in the moment.

3. Build consistent, multi-source coverage to shape parametric memory. The closed-book engine names brands from training memory alone. The only way to influence that channel is sustained, consistent, multi-source coverage over time, always-on rather than burst-shaped, so that what the next model "knows" about your category is accurate and includes you.

4. Restructure content answer-first, with schema. Once an engine browses, it lifts the cleanest, most current, most quotable page. Apply the answer-first specification and citation-relevant schema to cornerstone content and to the contributed and community content you can influence. This is the fastest lever, measurable in weeks, and in low-concentration long-list categories it is often what elevates you from mid-list mention to named leader.

5. Fix the knowledge graph: Wikipedia, knowledge panels, entity data. Wikipedia recurred as a cited source in our corpus, and it feeds both retrieval and parametric memory. A complete, well-sourced Wikipedia presence (where notability genuinely supports it), an accurate Google knowledge panel, and consistent structured entity data are foundational, and are exactly where Indian brands most often lag global peers.

6. Compete on prominence, not just inclusion. Because concentration is low (top-3 shares of 5.5 to 20%), being one of a dozen names is cheap and weak. Target ordinal position and framing: being named first, described as the category leader, or singled out in the answer's opening line, the part browsing engines lift most. In Healthcare and other YMYL categories, where concentration is genuinely high, this means fighting for durable top-three consensus.

7. Enforce consistent messaging and facts across the web. Because both retrieval and parametric memory reward consensus, inconsistent descriptions of your product, pricing, or positioning weaken citation and increase the risk of inaccurate closed-book assertions. Publish canonical facts and keep them consistent everywhere the brand appears.

8. Monitor and govern the closed-book band. Continuously query the closed-book and browsing engines alike, catalogue where your brand is asserted without a source, and correct the underlying corpus. In YMYL categories, where a third of answers were unsourced, treat this as compliance-grade monitoring, not optional hygiene.

Executed together, this playbook does not chase a single assistant or a single trick. It builds the coverage, structure, and consistency that the browsing engines cite and the closed-book engine remembers, which is the only durable way to move from present but unverifiable to cited, prominent, and correct.

Methodology and data

Design. This report is built on a real, primary audit conducted by Melivana's PR Intelligence unit, not a modeled estimate. On 8 July 2026, we posed 20 India-context "best/top X in India" questions, four per sector across Fintech, SaaS, D2C, Healthcare, and Manufacturing, to three live AI assistants with web search enabled: ChatGPT (gpt-4o), Google Gemini (2.5-flash), and Perplexity (sonar-pro). Each question was posed to each assistant, yielding 60 answers (20 questions × 3 assistants; 12 answers per sector). Across those answers we recorded 682 brand mentions and 334 citation instances spanning 237 distinct domains.

Engine behaviour (reported honestly). The three engines did not behave alike, and the difference drives several findings. Gemini and Perplexity browsed the live web and cited sources on 100% of their answers. ChatGPT (gpt-4o), despite web search being enabled, answered closed-book on 100% of its answers, producing zero citations while still naming, on average, roughly nine Indian brands per answer. Consequently, 40 of 60 answers (66.7%) carried a source, and 79.5% of all brand mentions occurred in answers that carried a source.

Representative questions. The bank used India-context framing throughout ("in India," "for Indian businesses," GST/UPI/Indian-market qualifiers where natural). Examples by sector:

• Fintech: best UPI apps for small businesses in India; top digital lending platforms; best neobanking apps for Indian startups; best expense-management fintech.
• SaaS: best accounting software for GST compliance in India; top HR/payroll software for Indian mid-market; best CRM for Indian small businesses; leading Indian customer-support automation tools.
• D2C: best Indian D2C skincare brands; top Indian D2C coffee brands; best Indian sustainable clothing brands; most-recommended Indian D2C nutrition brands.
• Healthcare: best online pharmacy apps in India; top telemedicine platforms; Indian diagnostics labs with home collection; best digital health apps for chronic-care management.
• Manufacturing: leading electronics contract manufacturers in India; top auto-component manufacturers; best industrial-automation companies; leading Indian solar-module manufacturers.

Coding scheme. Each answer was coded for: (a) whether any Indian brand was named (result: 100% to 0% named none); (b) whether the answer carried an identifiable source (inline citation or linked domain), the strict "sourced" standard behind the 66.7% headline; (c) which brands were named and how often, to compute per-sector averages, leaders, and top-3 concentration; (d) which domains were cited, feeding Report 7's citation index; and (e) unsourced (closed-book) mentions, feeding Finding 5. Per-sector brand averages ranged from 8.5 (Fintech) to 13.75 (D2C); the full mention total reconciles to 682 across the corpus.

Consistency across the trilogy. Reports 6, 7, and 8 draw on this same 8 July 2026 corpus, sectors, and query set, so their figures reconcile by construction: this report's citation and concentration metrics, Report 7's source-level citation index (the YouTube/Reddit/LinkedIn/Instagram/Wikipedia/aggregator landscape), and Report 8's share-of-answer league tables are three views of one dataset.

How to use this benchmark

This report is a diagnostic and a planning instrument, not a scorecard to file away. Use it in four ways.

As a baseline. Adopt the sourced-citation standard (named with a source) as your primary AEO KPI, not mere presence, which reads near-100% in India-context queries and hides the real gap. Instrument your own version of this audit and track sourced-citation rate, mention-sourcing share, top-3 concentration, and closed-book mentions over time.

As a prioritisation lens. Decide which problem you have. In flat, low-concentration categories (D2C, SaaS), presence is cheap and your priority is prominence, position and framing within a long list. In the concentrated YMYL exception (Healthcare), your priority is durable top-three consensus. Everywhere, a third of answers arrive unsourced, so governance is a constant.

As a business case. The coverage-to-citation link (Finding 3) and the answer-first structure lever (Finding 4) translate directly into investment logic: consistent, multi-source coverage on the surfaces engines actually cite, plus answer-first structured content, are the two highest-return AEO levers, and this real audit shows exactly why and where.

As a governance trigger. Treat the closed-book band (Finding 5) as a monitoring mandate, especially in YMYL categories, and build the ongoing, multi-engine audit cadence that lets you both grow and govern your AI representation.

Read alongside Reports 7 and 8, this benchmark completes a single argument: Indian brands are already present in India-context AI answers; the real contests are sourcing and concentration; the sources that decide sourcing are social and aggregator surfaces, not the press; and a third of answers arrive with no source at all, so the discipline that wins is not getting named, but being verifiably, prominently, and correctly represented.

Limitations

This is a real audit, and being candid about its bounds strengthens rather than weakens its credibility. Six limitations bear emphasis, and each points to how the audit should evolve as a repeatable, cadenced instrument rather than a one-off.

First, bounded sample. The study rests on 60 answers, 20 discovery-style questions, three assistants, a single date (8 July 2026). It is a genuine primary snapshot, but a snapshot: robust enough to reveal structure (engine behaviour, presence, concentration, sourcing) but not a census. Running the same design monthly, and widening the question bank, would tighten every figure.

Second, one model per engine. We queried one model per assistant, gpt-4o, Gemini 2.5-flash, Perplexity sonar-pro. Different models within the same product family behave differently, so results should not be read as fixed properties of "ChatGPT" or "Gemini" writ large.

Third, and most important, the closed-book condition reflects gpt-4o, not all of ChatGPT. gpt-4o did not browse in this run despite web search being enabled, producing the 33.3% unsourced band single-handedly. Other ChatGPT models and configurations do browse and cite; our finding is that this engine, in this run, answered closed-book on every question. Read it as a real and important behaviour observed, not as a universal claim about the product.

Fourth, query-type effect on concentration. "Best/top X in India" list queries deliberately elicit long brand lists, which mechanically produce low top-N concentration. Narrower or comparison queries ("X vs Y," "best UPI app for freelancers") would return shorter lists and higher concentration. The concentration figures (5.5 to 20% top-3) are properties of this query type, not of the categories in the abstract.

Fifth, judgment in tallying. Counting brands, classifying a name as an "Indian brand," and attributing mentions to sources all involve coding judgment at the margins. We applied consistent rules, but a different coder might shift individual counts slightly; the structural findings are robust to such variation, the third-decimal precision is not.

Sixth, scope. The audit covers five sectors and English-language, India-context queries. Vernacular-language querying, a fast-growing and strategically vital surface in India, is out of scope here and is a priority for future rounds.

None of these caveats changes the strategic conclusions, total presence, engine-driven sourcing, low concentration outside Healthcare, a social-and-aggregator citation landscape, and a large closed-book risk band. They bound the precision of the specific numbers and define the cadence on which this audit should be repeated. That repeatability is the point: a single run reveals the structure; a standing cadence turns it into an intelligence function.

Appendix / glossary

• AEO (Answer Engine Optimization): The discipline of getting a brand named, and ideally cited with a source, inside the answers AI assistants generate.
• GEO (Generative Engine Optimization): The broader practice of engineering content, structure, and reputation so generative AI systems surface a brand; often used interchangeably with AEO.
• AI assistant / answer engine: A system that returns a synthesised answer rather than a list of links, here, ChatGPT (gpt-4o), Google Gemini (2.5-flash), and Perplexity (sonar-pro).
• Sourced citation: An instance where an assistant names a brand and the answer carries an identifiable supporting source (inline citation or linked domain). The primary metric in this report, 66.7% of answers.
• Closed-book / unsourced mention: A brand named with no verifiable supporting source, generated from parametric memory, the entire 33.3% unsourced band, produced by gpt-4o in this run.
• Parametric knowledge: What a model "knows" from training data, without live retrieval; the sole channel behind closed-book answers.
• RAG (Retrieval-Augmented Generation): Grounding answers in live-retrieved sources; the channel Gemini and Perplexity used on 100% of answers.
• Consensus / Matthew effect: Models privilege claims agreed across independent reputable sources; relevant to why category leaders recurred even in long lists.
• YMYL ("Your Money or Your Life"): Trust-sensitive categories (finance, health) where models apply stricter filtering; Healthcare showed the highest concentration and lowest mention-sourcing here.
• Answer-first content: Content that leads with a direct, self-contained answer under a question-shaped heading, chunked for retrieval, with provenance and schema.
• Top-N / top-3 share: The proportion of in-category brand mentions captured by the most-mentioned N brands; a measure of concentration (5.5% D2C to 20.0% Healthcare).
• Mention-sourcing share: The proportion of all brand mentions that landed in sourced answers, 79.5% across the corpus.
• The trilogy: Reports 6 (this report, brand visibility/citation), 7 (India Media Citation Index, source-level), and 8 (AI Answer Share, competitive share), all analysing the one shared 8 July 2026 query round across the same five sectors.

Melivana | PR Intelligence Series 2026, Report 6 of 8. Prepared by the PR Intelligence unit. Figures are real counts from a primary audit of 60 AI answers across three assistants, conducted 8 July 2026; the audit is deliberately bounded (see Limitations) and designed to be repeated on a regular cadence.