What Google Says About AI Citation vs What Sells

A useful thing happened in May 2026. Google published its first proper guidance on optimising for generative AI features in Search, and a great deal of what the GEO industry had been selling for two years was contradicted in writing by the company whose product it claimed to optimise for.

This matters if you are buying AI visibility work in India, because the gap between what Google says is required and what vendors say is required is now documented rather than debatable. You can hold a proposal up against the guidance and check. That is unusual in this category and worth using.

What follows sets Google's stated position beside the claims most commonly made in sales conversations. Some of the popular tactics are harmless but oversold. A couple are directly contradicted. And one genuinely useful development arrived in June 2026 that a surprising number of vendors have not caught up with.

The schema conversation, settled

Structured data has been sold as an AI-citation lever for two years. The claim usually runs that adding enough markup makes a page machine-readable, and machine-readable pages get cited. It sounds mechanical and therefore credible.

Google's position is narrower and clearer. Structured data helps Google understand a page and can make it eligible for rich results. There is no special markup for the generative features, and the guidance explicitly names overfocusing on structured data as a mistake. Those two statements can both be true at once, and they are: accurate markup is worth having, and it is not the mechanism people are being sold.

Keep the markup. Keep it accurate, and keep it consistent with what a human sees on the page, because markup that describes content the page does not contain is a quality problem rather than a shortcut. Just stop paying a premium for it as a citation guarantee, and be suspicious of any proposal whose central deliverable is schema volume.

The AI directive file

The same applies to the text-file convention that circulates as an AI best practice. Google states that you do not need to create new machine-readable files, AI text files, markup or alternative document formats for its generative features.

That is a direct answer to a widely sold deliverable. It does not mean the file causes harm, and other systems may choose to read such files. It means Google, specifically, does not require it, and a proposal that leads with it is leading with something the platform has publicly declined to endorse. Ask what the second deliverable is.

Common sales claim	Google's documented position	How to respond
Deep schema markup gets you cited	No special schema.org markup exists for AI features. Overfocusing on structured data is named as a mistake	Keep markup accurate. Refuse to pay for it as a citation lever
You need an AI directive text file	Google states you do not need to create new machine-readable files, AI text files or markup	Ask what else is in the scope of work
We can guarantee citation in a named assistant	Meeting all stated requirements does not guarantee crawling, indexing or serving	Treat a guarantee as a disqualifier
Our proprietary score measures AI visibility	Google advises being wary of third-party tools that promise ranking success or claim to use internal Google metrics	Ask for raw responses, cited URLs and a dated run log
AI visibility cannot be measured in Google's tools	A Generative AI performance report was announced for Search Console on 3 June 2026, rolling out to a subset of sites	Check your own property before accepting the claim
Ranking first means you will be cited	Indexing and snippet eligibility are the stated gates. Position is not stated as the selection mechanism	Ask for citation data, not a rankings chart

What changed in June 2026

For most of the past two years, one honest objection to AI visibility work was that Google's own tooling did not separate it out. On 3 June 2026 Google announced a Generative AI performance report in Search Console, rolling out to a subset of sites.

Two consequences follow. The first is that anyone still telling you AI performance is unmeasurable in Google's tooling is working from stale information, and you should check your own property rather than take their word for it. The second is subtler and more important: a report about performance is not a report about citation. Being cited inside an answer frequently produces no click at all, so a traffic-shaped report and a citation-shaped report measure different things and both are needed.

Keep them as separate layers. Visibility, meaning whether you are named and cited. Traffic, meaning whether anybody arrived. Collapsing the two into a single number is how programmes end up looking successful while pipeline stays flat.

The one India measurement worth knowing, and how to read it

Most India-level AI figures in circulation are unusable, for reasons covered above. There is one measurement of Indian brand visibility that is worth having, and it comes with a caveat you should apply before quoting it.

An India-focused analysis published for the first quarter of the 2026-27 financial year examined 2,981 brands across 59,620 prompts between April and June 2026. It found that 19.3 percent of those brands met its high AI visibility benchmark, leaving 80.7 percent below it, with an average visibility score around 30 out of 100.

The caveat is that this is a single-vendor benchmark using a self-defined scoring method, not an independently audited industry statistic. No standards body validates the threshold, and a different scoring design would produce a different split. Treat it as one credible, disclosed, India-specific data point rather than as a market truth.

Read that way it is still genuinely useful, because it is the only India-specific brand-visibility measurement located across four independent research passes. It says the field is wide open rather than crowded. Most Indian brands are not competing for citation because most Indian brands have not started, which is a very different market condition from a saturated one and calls for a different level of investment.

Adoption figures, and the split people miss

Two very different numbers circulate about how many Indians use generative AI, and both are correct because they measure different populations.

On the usage side, the platform itself disclosed 100 million weekly active ChatGPT users in India in February 2026. That is a first-party figure and it is large. On the diffusion side, the Microsoft AI Economy Institute's global adoption work measured India's population-level diffusion rate, meaning the share of the population estimated to be using generative AI tools, at 14.2 percent in the first half of 2025 rising to 15.7 percent in the second half, against a global average around 16.3 percent.

Those are not in conflict. A very large absolute user base and a population share slightly below the global average are both true of a country with India's population. The error to avoid is quoting one as though it settled the other, which happens routinely in proposals where the large number appears and the diffusion number does not.

For a Pune exporter, neither number is the operative one anyway. Your buyer is in Singapore, Frankfurt or Chicago, and how many Indians use an assistant tells you nothing about whether that buyer's assistant names you. Domestic adoption statistics are context. Your destination market's behaviour is the measurement, and it has to be produced rather than cited.

The one thing Google will not tell you

Everything above concerns Google. It covers AI Overviews and AI Mode, and it is authoritative for those surfaces. Your buyers are not only using those surfaces.

Assistants retrieve differently from one another, and they draw on different pools of material. The same question about your category can be answered largely from business press in one assistant and largely from trade and technical material in another. Google's guidance says nothing about this, correctly, because it is not Google's product. It means a programme built only against Google's documentation covers part of the surface area your buyer actually touches.

This is where measurement stops being optional. There is no published document telling you which sources a given assistant favours for your category in your destination market. The only way to know is to run the prompts and code the citations. Anybody presenting a general rule here, of the form assistant X prefers source type Y, is generalising from someone else's prompt set at some earlier moment.

Volatility compounds it. Citation share for a given source type has been observed shifting sharply within weeks as retrieval behaviour changes. A single favourable capture is a snapshot of one moment, not a description of a market. Repeat the same prompts in clean sessions, record the run details, and report the spread rather than the best result.

Two questions that have not been answered by anyone

Honest gaps are worth naming, because vendors fill them with confident-sounding claims and buyers have no way to check.

The first is whether location changes which sources an assistant draws on. It seems plausible that the same prompt asked from Pune and from London would retrieve differently, and plenty of proposals assert it. No public controlled study located in this research holds prompt intent, model version, session state and account history constant while varying only location and language, then codes every cited URL by country of origin. Until one exists, the claim is a hypothesis. Testable against your own category, cheaply, but a hypothesis.

The second is whether Indian-language pages get cited in Indian-language answers. Retrieval in Indic languages is a documented engineering problem in its own right, and content in those languages is thinly represented in training material. What nobody has published is a measurement of whether a well-structured Marathi page actually earns the citation, or whether an English source carries it while the answer is delivered in Marathi. Those are different outcomes with different implications for where budget goes.

Neither gap is a reason to avoid the work. Both are reasons to run the work as a test with a decision point rather than as a settled tactic with a fixed deliverable. And both are questions worth putting to any vendor, because the answer tells you quickly whether you are talking to somebody who reads primary sources.

What is worth paying for instead

If schema volume and directive files are not the product, something has to be. Three things are, and none of them is glamorous.

The first is a fixed, disclosed prompt panel written from your actual buying situations, with a baseline captured before anything changes. This is the closest thing the category has to a contract. It defines what is being measured, it prevents the goalposts moving, and it is the artefact you should ask to see in the first meeting rather than the fifth.

The second is source-side work. The majority of citations in a generated answer point somewhere other than the brand's own website, which means third-party presence is infrastructure rather than a separate public relations line item. For an exporter, that means trade publications, industry associations, standards bodies and reference material in the destination market, not in aggregate.

The third is entity consistency. One coherent set of facts about your company across your own site, company records, professional profiles and industry directories. Where those disagree, systems hedge or reconcile badly, and you inherit whichever version they picked. This is unfashionable, cheap and effective.

Reading a proposal in ten minutes

Take any AI visibility proposal and look for four things. Is there a named prompt panel, or only a promise of one. Is there a baseline date. Are mention, citation and recommendation defined separately, or folded into a single score. And is there any guaranteed outcome anywhere in the document.

If the panel is missing, you are buying a report you cannot verify. If mention and citation are merged, you will never know which one moved. If a guarantee appears, the proposal contradicts the platform's own documentation, which is a reasonable place to stop.

None of this requires technical expertise. It requires reading the guidance once and holding the proposal beside it. Google published the guidance for site owners, not only for specialists, and using it as a purchasing checklist is entirely within its intended purpose.

---

Frequently Asked Questions

Does structured data get a page cited by AI?

No. Google's guidance states there is no special schema.org markup for its generative features, and it specifically names overfocusing on structured data as a mistake. What structured data does is help Google understand a page and support rich-result eligibility, which is worth having on its own terms. Keep the markup, keep it accurate and consistent with what a human sees on the page, and stop treating it as a citation mechanism.

Do we need an AI directive text file on our site?

Not for Google. Its guidance states that you do not need to create new machine-readable files, AI text files, markup or alternative document formats for its generative features. Other systems may choose to read such files, so the file is not harmful, but a proposal whose headline deliverable is that file is leading with something Google has publicly declined to endorse.

What does Google actually require for a page to appear in AI Overviews or AI Mode?

Two things: the page must be indexed, and it must be eligible to be shown with a snippet. Google states there are no additional technical requirements beyond meeting Search's standard technical requirements. It also states plainly that satisfying every requirement does not mean a page will be crawled, indexed or served, which is why guarantees in this category are not credible.

Can AI visibility be measured inside Google's own tools?

Partly, and this changed recently. Google announced a Generative AI performance report in Search Console on 3 June 2026, rolling out to a subset of sites. Anyone telling you AI performance is unmeasurable in Google tooling is working from outdated information. That said, a performance report measures traffic-side outcomes, and being cited inside an answer frequently produces no click at all, so citation still has to be measured separately.

How do we tell a serious GEO vendor from a repackaged one?

Ask for the prompt panel and the baseline date in the first meeting. A serious programme has a fixed, disclosed set of buyer-intent prompts, a baseline captured before any intervention, and separate definitions for mention, citation and recommendation. If those are missing, or if any outcome is guaranteed, you are looking at a report you cannot verify.

Should a proprietary visibility score be treated as evidence?

Only if you can see what is underneath it. Google itself advises being wary of third-party tools that promise ranking success or claim to use internal Google metrics. A score is acceptable as a summary of raw data you are also given. It is not acceptable as a substitute for the raw responses, the cited URLs and a dated run log.

Is there a reliable figure for how much traffic Indian publishers lost to AI answers?

No. No credible study located in this research isolates Indian publishers and measures traffic change causally attributable to AI answers. The widely quoted decline figures were measured in other markets. Presenting them as Indian numbers is a fact-merging error, and a vendor doing it has either not checked the provenance or is relying on you not checking.

If schema and directive files are not the product, what is?

Three things. A fixed, disclosed prompt panel with a dated baseline, which is what makes any later claim verifiable. Third-party authority in your destination markets, since most citations in a generated answer point somewhere other than the brand's own site. And entity consistency, meaning one coherent set of facts about your company across your site, company records, professional profiles and industry directories, so systems are not reconciling several versions of your own description.

---