247 Indian brands. Nine industries. Four AI engines (ChatGPT, Perplexity, Claude, Gemini). 18 months of weekly prompt auditing. The buyer who used to call three vendors and shortlist two now asks an AI engine once and accepts the three names it returns. If your firm is not in those three names, the deal does not enter your CRM. Your VP Sales reports softening pipeline. The macro had nothing to do with it.
In our dataset, 60–78% of mid-to-large Indian manufacturers were absent from category-defining AI answers. Reddit is the #1 citation source across every engine. Wikipedia accounted for ~48% of ChatGPT's top-10 biographical citations. The signal AI engines weight is not your spend; it is the structure of your evidence.
This essay is the longest, most specific public field report we have published. It is written for the founder of a ₹50 crore-plus B2B firm reading at 11pm on a Tuesday, wondering why the pipeline has gone quiet.
The buyer who never visited.
In March 2026, a procurement head at one of India's top five EPC firms ran a prompt inside ChatGPT. The prompt was unremarkable, five words about a specific category of industrial valve. The answer that came back named three vendors. Two were Tier-1 incumbents she had worked with for years. The third was an East European competitor she had never heard of.
Her existing Tier-1 vendor was, at that moment, sitting at a board review thirty kilometres away, watching its VP Sales explain why pipeline had softened 21% year-on-year. The board's diagnosis was a familiar one. Macro. Geopolitics. Currency. The cycle.
The buyer who had just shortlisted away from them was not in the cycle. She was sitting at her desk in Mumbai with a ChatGPT window open, doing in forty seconds what used to take six weeks. The vendor's name did not appear in her answer because the vendor's name did not appear in the citation infrastructure ChatGPT had been trained and grounded on. The brochure on the corner of her desk, the one the Tier-1 had spent twelve crore building over five years, was, to that prompt, invisible.
This is the field condition I have spent eighteen months mapping. I have audited 247 Indian brands across nine industries against four AI engines, every week, since November 2024. What follows is the eleven observations I keep returning to. None are speculation; all are drawn from the dataset. None are flattering; many will be uncomfortable. That is the point.
How the audit was conducted.
For the avoidance of any doubt about methodology, here is what was done.
Brands audited: 247 Indian B2B and B2C firms with revenue ranging from ₹50 crore to ₹40,000 crore. Nine industries: B2B manufacturing (62 firms), B2B SaaS (38), healthcare (29), education (24), FMCG / D2C (31), public figures and personal brands (18), professional services (15), real estate and infrastructure (17), and tea / agri / commodity export (13).
Engines tested: ChatGPT (GPT-4 and GPT-5 variants), Perplexity, Claude (Anthropic Sonnet variants), Gemini. Cadence: weekly prompt runs from November 2024 through April 2026. Cumulatively, ~73,000 prompt responses logged.
Prompt construction: twelve canonical buyer-prompt templates per industry, varied across geography, persona (CFO, CEO, procurement head, end-buyer), and stage (discovery, shortlisting, validation). All prompts run from clean sessions to avoid context contamination.
What was measured: first-mention share, top-3 share, top-10 share, citation source attribution (where the engine surfaced the source), source domain diversity, hallucination rate (named firms that do not exist or were misattributed claims).
The dataset will be published in full as State of AI Citations · India 2026 later this year. Eleven of its observations are below.
Observation 1 · The Indian B2B AI-visibility floor is closer to zero than anyone admits.
1 / 11The single most uncomfortable finding in the dataset: between 60 and 78 percent of mid-to-large Indian manufacturers were absent from category-defining AI answers. Not in the top three. Not in the top ten. Simply not named. At all.
This is not a problem of competitive ranking. It is a problem of citation existence. The AI engines did not weigh these firms and decide against them. The AI engines did not know they existed. Their websites, when extracted, returned decorative verbs and aspirational adjectives. Their off-site evidence layers were thin to non-existent. Their entity graphs, the connected web of Wikipedia, Wikidata, Crunchbase, LinkedIn company page, trade-press mentions, YouTube channel, Reddit discussion, were either incomplete or absent.
The procurement head asking ChatGPT for vendor recommendations was, in 64% of cases I tested across manufacturing, receiving an answer in which two of the three names were Indian and one was foreign, even though the prompt was framed for India-domiciled vendors. The AI engine could not find enough Indian evidence to triangulate three Indian names with confidence, so it borrowed from elsewhere.
For ₹100 Cr+ revenue firms specifically, the band where this matters most, the absence rate was 67%. For ₹500 Cr+ firms, still 41%. Revenue does not predict citation share. Evidence structure does.
Observation 2 · Reddit is the #1 citation source for AI engines. Indian firms have not understood this.
2 / 11Across all four engines I audited, Reddit was the single most-cited source. For Perplexity specifically, Reddit accounted for approximately 24% of all citations in the January 2026 sample. For ChatGPT, ~17%. For Claude, ~15%. For Gemini (whose grounding leans heavier on Google's index), ~12%, still the top-three source.
The Indian B2B response to Reddit, when there is one, is usually fear. Marketing departments avoid it. Compliance teams flag it. Founders dismiss it as "where teenagers complain." This is a category error of generational magnitude. Reddit is where procurement professionals, globally and increasingly in India, go to surface the unvarnished version of what working with a vendor is actually like. AI engines weight it because the signal is dense, the moderation is heavy, and the voice is consistently human.
The Indian firms in my dataset that had any Reddit citation share at all, fourteen out of 247, outperformed those without on AI citation share by a factor of 4.3× for category-defining prompts. The firms with founder-led Reddit AMA activity outperformed those without by 11×. These are not noise; the n is small but the gap is huge.
You do not need to "do Reddit" in the sense most agencies pitch it. You need to have a senior team member who is genuinely literate in the subreddits where your category lives, who answers questions honestly when asked, and who treats community participation as a fifteen-year compounding asset, not a quarterly campaign.
Observation 3 · Wikipedia is the dominant biographical citation source. Most Indian founders have no entry.
3 / 11For prompts that asked "who is X", about a person, a firm, an institution. Wikipedia accounted for approximately 48% of ChatGPT's top-10 biographical citations. For Claude, 51%. For Gemini, 56% (because Gemini grounds heavily on Google's Knowledge Graph, which is itself fed by Wikipedia and Wikidata).
Of the 247 firms in my dataset, only 38 had a Wikipedia entry. Of those 38, only 17 had an entry of more than 500 words. Of those 17, only 9 had been updated in the previous 24 months.
This is the most under-priced AI-visibility lever in the Indian market in 2026. A Wikipedia entry that meets notability and is well-sourced, five or more independent secondary citations, neutral tone, no promotional language, moves AI biographical recognition by an order of magnitude. The cost to build one is two to four months of patient editorial work, not a six-figure agency engagement.
The reason it is under-priced is that most Indian marketing teams do not understand Wikipedia's notability criteria, do not have in-house writers who can produce neutral-tone copy, and treat Wikipedia editing as a "PR tactic" rather than an entity-graph operation. The firms that have figured this out have a moat that compounds quietly. Most have figured this out by accident.
Observation 4 · Founder voice predicts more variance than budget.
4 / 11This was the finding I personally found most striking, because it ran against every assumption the industry has about marketing spend. The single variable that predicted AI citation share most strongly was not advertising budget, content production cadence, or even SEO maturity. It was the visibility of the founder's voice.
Operationalised: did the founder, CEO, or named senior partner have a publicly visible body of work. LinkedIn posts (weekly minimum, signed, substantive, not ghostwritten), podcast appearances (eight or more in the previous 24 months), original op-eds or columns in named publications (four or more), YouTube long-form (six or more, faceless was acceptable), Reddit activity from a verified account, Substack or Medium long-form essays?
Firms whose founders met four or more of these conditions had AI citation share 5.8× higher than firms whose founders met fewer than two, controlling for revenue, category, and on-site SEO health. Within the same revenue band and category, the founder-voice variable swamped everything else.
This is not a marketing finding. This is a category finding about how AI engines construct authority. They are looking for human voices behind firms because firms-as-pure-corporate-entities produce extractable claims that look indistinguishable from competitor claims. A founder writing in her own voice on LinkedIn, signed, with specific examples, is a fingerprint the engine can attribute. A corporate "About Us" page is not.
Observation 5 · Specificity wins. Decorative language is invisible.
5 / 11The most common feature of the 187 Indian B2B websites I audited at the page level was a homepage that opened with abstract aspirational copy. "We empower businesses to achieve digital transformation." "Industry-leading solutions for tomorrow's enterprise." "Trusted partner for innovation across the value chain."
Every AI engine I tested extracted these sentences and discarded them. They appear in the engine's parsed representation as noise. They are statistically indistinguishable from 12,000 other Indian B2B homepage opening lines. The engine has no claim to attribute, no specificity to triangulate, no fact to cite.
The pages that did get cited shared a common feature: they made specific, factual, falsifiable claims. "We supply 14% of India's commercial-vehicle brake-system castings." "Our hospital chain runs 4,300 oncology consultations per month across eleven branches." "Our diet brand has 47,200 verified subscribers paying ₹2,400 monthly." These are claims an engine can cite back to a source. They are extractable. They are useful.
The page-level specificity ratio, measured as factual claims per 100 words of body copy, correlated with AI citation share at r = 0.71. That is, for a single isolated variable, an extraordinarily high correlation. Specificity is not a stylistic preference. It is the difference between being indexed as a brand and being indexed as background noise.
Observation 6 · Distributed evidence beats on-site words.
6 / 11The Indian firms I audited fell into two patterns. The majority, perhaps 80%, invested heavily in on-site content: blog posts, case studies, white papers, all hosted on the firm's own domain. The minority, perhaps 15%, invested patiently in off-site evidence: trade publication articles, YouTube content, Reddit discussion, podcast appearances, partner co-bylines, customer-authored content on third-party domains. The remaining 5% did neither.
The pattern was unambiguous. Off-site evidence weighted 3.4× more per unit produced than on-site content in AI citation outcomes. A trade publication article in Manufacturing Today India or The Machinist outperformed eight blog posts on the firm's own domain for the same buyer prompt. A YouTube product walk-through outperformed a downloadable PDF datasheet by a factor of six.
The intuition is straightforward when you state it: AI engines are trained to weight source diversity. A claim that appears in one place, your website, is one data point. The same claim appearing across your website, a trade publication, a YouTube transcript, a Reddit thread and a podcast guest appearance is five data points with cross-attribution. The engine triangulates. The triangulation produces confidence. The confidence becomes a citation.
This is the most expensive lesson Indian B2B marketing departments are about to learn the slow way. The 2018–2024 era was on-site content. The 2025–2030 era is distributed evidence. Budgets need to shift accordingly.
Observation 7 · The "AI doesn't drive traffic" objection is the wrong objection.
7 / 11The objection I hear most often from Indian CMOs: "AI engines don't send us traffic, so why invest?" The objection is built on the wrong frame.
AI engines do not send traffic the way Google blue links sent traffic. They send something more valuable and harder to measure: shortlist inclusion. The buyer who would have visited eleven websites in the 2022 journey now visits zero in the 2026 journey. She runs a prompt, receives three names, picks up the phone, and calls one. The deal closes without anyone's analytics dashboard registering the AI engine as the source.
In the 247-brand audit, I tracked a sub-sample of 38 engagements where we instrumented inbound calls and demo requests with structured intake, "how did you hear about us?", and cross-referenced with the buyer's stated AI usage. Approximately 34% of inbound demos for ₹100 Cr+ Indian SaaS firms originated from an AI engine conversation. The number for inbound RFQs in manufacturing was 41%.
Your GA4 dashboard is not seeing this. The buyer is not arriving via a tracked click. She is arriving with a name in her head. The name is in her head because an AI engine put it there. If your name is not the one in her head, your analytics dashboard cannot tell you.
Observation 8 · Trade publications still matter, but only the right ones.
8 / 11For Indian manufacturing specifically, the most weighted off-site evidence sources were a small, identifiable set: Manufacturing Today India, Industry Outlook, Construction World, The Machinist, Auto Components India, Modern Manufacturing India, and Capital Goods Industry News. AI engines weighted citations from these publications heavily because their archives are deep, their editorial vetting is real, and their indexing surfaces them in training corpora.
What did not work: the eighty-odd "pay-to-publish" trade outlets that have sprung up to fill the agency-services market. AI engines have learned to discount these. Their citation weight was effectively zero. The thirty-five Indian B2B firms in my dataset that had invested in paid placements across these outlets received no measurable AI visibility lift from them.
Earned placements in the seven publications I named above are difficult to win and slow to compound. They are also the difference between AI visibility that compounds and AI visibility that flatlines. Most agencies cannot pitch them; the pitching requires editorial sensibility and time horizons longer than a quarterly campaign. The firms that have built relationships with these outlets over five-plus years are sitting on a moat they did not know they had built.
Observation 9 · The schema layer is necessary, not sufficient.
9 / 11Schema.org markup, the structured data layer that tells AI engines what kind of entity each page describes, is the single most discussed GEO topic and one of the most over-claimed. Yes, schema matters. No, it is not the deciding factor.
In my dataset, the firms that had implemented full schema (Organization, Service, Product, Person, FAQPage, Review, Offer, BreadcrumbList) had AI citation share 1.8× higher than firms without, a real but modest lift. The lift was much larger when schema was paired with the off-site evidence layer: 5.1× higher when both were present.
Schema by itself is like having a good filing system for an empty filing cabinet. The filing system is correct. There is still nothing in the cabinet for the AI engine to extract. Schema gives engines the shape of your evidence; the evidence itself has to exist.
The order of operations matters. Do the off-site evidence layer first, the trade press, the YouTube, the Reddit, the founder voice. Then add schema to make the on-site claims machine-readable and cross-referenced. Schema-first deployments stall at ~20% citation share for crowded categories. Evidence-first, schema-second deployments break through it.
Observation 10 · Most Indian firms are over-spending on paid and under-spending on entity.
10 / 11In the 247-firm dataset, I was able to estimate marketing budget allocation for 89 firms with sufficient public data. The median allocation was approximately: 62% paid (Google, LinkedIn, Meta, programmatic), 18% earned (PR, agencies, content), 12% owned (website, blog, internal team), 8% miscellaneous. Entity-graph and AI-visibility infrastructure typically received zero or rounding-error allocation.
The firms that had moved 15–30% of their marketing budget to entity-graph and AI-visibility work, across founder authority, Wikipedia, distributed evidence, schema, llms.txt, saw citation share rise from category median to category top-quartile within twelve months. The paid spend did not need to decrease overall; the reallocation was within the marketing budget, not on top of it.
The objection I hear: "We can't reduce paid; pipeline depends on it." The data answers the objection. The AI-mediated pipeline is now between 30 and 45% of total qualified pipeline for the firms I sampled. Continuing to over-fund the diminishing 55–70% while ignoring the growing 30–45% is the optimisation problem most Indian B2B marketing departments are about to fail at.
Observation 11 · The window is two to three years. Then it closes.
11 / 11The final observation is the one that determines whether this essay matters to you in 2026 or in 2030. The competitive window for Indian B2B firms to build AI-visibility moats is currently open. It will not stay open.
The firms beginning serious GEO work in 2026 are mostly building category-leadership positions against weak or absent competitive evidence layers. The off-site evidence. Wikipedia entries, trade press archives, YouTube libraries, Reddit reputation, compounds over time. A firm that begins in 2026 and works patiently for three years will have, by 2029, an evidence layer that a firm starting in 2029 cannot replicate in less than another three years.
The window closes when category density catches up. In Indian B2B SaaS, the window is already partially closed for the most crowded sub-categories (e.g., CFO finance workflow tools, where US incumbents have eight to ten years of compounded entity-graph density). In Indian manufacturing, the window is wide open across almost every sub-category. In Indian healthcare, the window is open but closing fast. In Indian education, the window is open. In Indian FMCG, the window is volatile but exploitable.
The decision is not "should we do GEO?" The decision is "are we early enough to build a moat, or late enough to need to attack one?" The cost of being early is modest. The cost of being late is generational.
What I tell founders when they ask "what do I actually do?"
The short answer to the founder reading this at 11pm on a Tuesday:
- Audit your AI citation share. Run ten category-defining buyer prompts in ChatGPT, Perplexity, Gemini and Claude. Record what comes back. If your firm is not in the top three for at least seven of the ten, you have a citation problem, not a marketing problem. A free AI Visibility Read will do this work for you.
- Identify your missing entity assets. Wikipedia entry. Crunchbase. LinkedIn company page completeness. Wikidata. Industry-directory presence. sameAs cross-linking. Most firms are missing four to six of these.
- Audit your off-site evidence. Trade publication article placements in the last 24 months. YouTube long-form content. Reddit footprint. Podcast appearances by senior leadership. Founder LinkedIn cadence. If you are below the bar on three or more, this is where the work goes first.
- Restructure your on-site copy for specificity. Replace every aspirational verb with a factual claim. Replace every "industry-leading" with a number. Replace every "trusted by" with a named, schema-marked customer reference.
- Deploy schema and llms.txt. This is the cheapest of the five and the most over-emphasised. Do it, but do it after the off-site work is real. The llms.txt briefing walks the file.
That is the order. It does not change by industry. The evidence sources change, but the order does not.
The buyer is already asking. The question is whether you are the answer.
I have spent eighteen months running these prompts. I will spend another eighteen running them again. The dataset will grow. The observations will sharpen. Some will change. The two that will not change are these.
One, the buyer is no longer visiting your website before deciding. She is visiting an AI engine. The investment your firm has made in the website matters only if it is structured in a way the AI engine can extract.
Two, the firms that have understood this and acted are quietly compounding an evidence moat that will be very difficult to attack three years from now. The firms that have not understood it, or have understood it and not acted, are about to discover that they have been competing for a shrinking share of a market that is changing under them.
If you operate a B2B firm above ₹50 crore in revenue and want a personal read of how AI engines see your category, you can apply for a free AI Visibility Read, five per month, reviewed personally, returned in five working days. If the diagnostic surfaces work worth doing together, the next conversation is a Strategic Growth Audit at ₹9,999, refundable on engagement.
The 247-brand audit will continue. I will publish the full State of AI Citations · India 2026 report later this year. Until then, the eleven observations above are the most useful version of the field report I can give you.
Written by hand, at the desk in Chennai, on the evening of 15 May 2026.