What do the operators actually publish?
Less than the market assumes, and the gap between the three major operators is instructive because it shows what transparency looks like when a company chooses it.
Google publishes eligibility conditions for its own AI surfaces: a page must be indexed and eligible to appear with a snippet, and there are, in its words, no additional requirements or special optimizations. It does not publish how it chooses among the eligible.
OpenAI publishes its crawlers and how to control them — GPTBot for training, OAI-SearchBot for search and citations, ChatGPT-User for user-initiated fetches — with IP ranges so operators can verify them. It does not publish how cited sources are selected.
Microsoft goes furthest, and in a different direction entirely: rather than explaining the mechanism, it shows publishers the outcome. That is a meaningful distinction, and it is the closest anyone has come to letting a site see its own position in AI answers.
It is worth being clear why this matters commercially. Where criteria are unpublished, no one can demonstrate compliance with them, and a market for compliance appears anyway — priced against rules nobody has seen.
What does Microsoft publish that others do not?
Per-page citation data. The AI Performance report in Bing Webmaster Tools reports “Total Citations” and “Average Cited Pages” for a verified site, along with “page-level citation activity” showing “citation counts for specific URLs”.
The most useful element is the grounding queries: “key phrases the AI used when retrieving content”. This exposes something no other operator reveals — not the question a user asked, but the query the system generated internally when it went looking for material. That is a view into the retrieval step itself.
Coverage extends across “Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations”. The partners are not named, which matters because it means the reported total aggregates surfaces a publisher cannot enumerate.
For anyone assessing whether coverage reaches AI answers, this is the only first-party instrument that exists. It applies to a site you control and verify, which means it measures your own pages rather than a placement on somebody else's domain — a real limitation for assessing bought coverage, and not one Microsoft could reasonably fix.
Documented disclosures as of September 2026
Source: Microsoft, Bing Webmaster Tools AI Performance announcement; Google Search Central AI features documentation; OpenAI crawler documentation.
Values rank the depth of disclosure, not quality of the product. No operator publishes selection criteria.
What does Microsoft still decline to explain?
The mechanism, and it says so by what it offers instead. Microsoft does not specify what makes content likely to be used for grounding. In place of an explanation it gives recommendations: strengthen expertise, improve structure with headings and clear question-and-answer formatting, support claims with evidence, keep material fresh, and stay consistent across formats.
Those are reasonable editorial suggestions and they are not a specification. Nothing there tells a publisher why one page was chosen over another for a particular answer, and Microsoft does not claim it does.
More striking is the warning attached to the data itself. Microsoft states the metrics reflect “how often your content is referenced” without indicating “ranking, authority, or the role of any page within an individual answer”.
That caution is aimed squarely at the obvious misuse. A citation count is not a score, not a measure of standing, and not evidence that a page carried the answer rather than being one of several sources brushed against. Any vendor converting these counts into an AI visibility score is building a metric its own data source has disclaimed — the same move that produced domain authority and topical authority, now applied to a fresher number.
What can be inferred, and what cannot?
Inference is available at the level of preconditions, and nowhere further.
A document must be reachable to be retrieved, so crawler access is necessary. It must be indexed where indexing governs eligibility, as Google states for its surfaces. It must contain material that answers a question, because retrieval systems search with queries. These are structural facts and they follow from how the systems are described.
What cannot be inferred is anything quantitative. No published data supports a claim that a particular format, length, markup pattern or publication tier raises citation probability by any amount. The operators do not publish selection criteria, so nobody outside them can calibrate against it, and every specific figure in circulation was produced by someone observing outputs and guessing at causes.
The grounding queries Microsoft exposes narrow this slightly, because they show what the system was looking for when it found you. Used carefully that is a genuine feedback signal about which questions your material answers — which is closer to editorial insight than to an optimisation lever, and is more valuable for being so.
Observational studies of AI answers do exist, counting which domains appear across sampled questions. Those measure outputs rather than criteria, and a domain appearing often may be doing so because it is large, old and widely linked rather than because of anything a publisher could replicate.
Why does transparency vary so much?
Because the operators have different relationships with publishers, and disclosure follows commercial interest rather than principle.
Microsoft is competing for search share and courting publishers, and giving site owners visibility into citation is a way to make its surfaces worth optimising for. Google, which holds the dominant position, has less to gain by exposing mechanism and more to lose, and its documentation accordingly explains eligibility and stops. OpenAI operates no webmaster relationship at all, so it publishes crawler controls — the minimum needed for site operators to consent or refuse — and nothing about selection.
None of this is improper. Ranking and retrieval systems are routinely kept opaque to limit manipulation, which is the same reason Apple gives for withholding its chart algorithm, as documented elsewhere in this series.
The consequence for buyers is what matters. Where operators do not publish criteria, nobody can sell compliance with them, and the varying levels of disclosure mean a claim true of one system may be baseless about another.
Disclosure can also be withdrawn. A reporting feature offered during a public preview is not a commitment, and a publisher building a measurement practice on one operator's dashboard is depending on a product decision rather than a standard.
What this means for anyone buying placement
Connect Bing Webmaster Tools for your own site, because the AI Performance report is free, first-party and the only direct evidence available about whether your material is being used in AI answers.
Read its numbers as Microsoft asks you to. Citations count references, not standing, and a rising count is a fact about frequency rather than proof of authority. Treat any product converting those counts into a score as a metric its source has already disclaimed.
Remember what it cannot see. The report covers sites you verify, so a placement on a publisher's domain will not appear in your dashboard at all — the citation, if it happens, accrues to them. This is worth knowing before buying coverage on the promise of AI visibility.
Media Morpheus sells placement and search visibility work. What can be honestly promised is material that is reachable, indexed and answers a specific question, which is the documented precondition for being retrieved by any of these systems. Selection among eligible documents belongs to the operators, none of whom publishes how it works, and nobody selling you coverage has access to it.
If a proposal quotes a citation-rate improvement, ask which system it applies to and where that system documents the figure. Microsoft is the only operator publishing citation data at all, and it disclaims the interpretation such a figure would require.
Methodology and limitations
This report relies on three sets of operator documentation: Microsoft's announcement of the AI Performance report in Bing Webmaster Tools, Google Search Central's documentation on AI features in Search, and OpenAI's published crawler documentation. Five limits apply. First, all three describe their own products, and disclosures change without notice; this describes the position at the date of publication. Second, no operator publishes how sources are selected for citation, so nothing here explains why any particular page was used. Third, Microsoft's coverage includes unnamed partner integrations, so the surfaces represented in its citation totals cannot be fully enumerated by a publisher. Fourth, we conducted no measurement of citation behaviour for this report and reproduce no third-party citation statistics, because the vendor studies located do not disclose methodology sufficient to assess them. Fifth, Bing's general webmaster guidelines were sought as an additional primary source but could not be retrieved in usable form, and no finding rests on them.
Conclusion
Asking what signals answer engines use assumes the operators have said, and largely they have not. Google publishes what makes a page eligible and stops before selection. OpenAI publishes how to admit or refuse its crawlers and nothing about what happens afterwards. Microsoft goes furthest by a distance, showing publishers which of their own pages were cited and even the internal queries used to find them — then cautioning that those counts indicate nothing about ranking or authority. That warning is the most useful sentence available on this subject, because the market's next move is always to convert a countable number into a score. The preconditions are knowable and worth meeting. The mechanism is not published by anyone, and cannot be sold.
Frequently Asked Questions
Can I see whether AI answers cite my site?
For Microsoft's surfaces, yes. The AI Performance report in Bing Webmaster Tools shows total citations, page-level citation activity and the grounding queries used to retrieve your content.
Does a high citation count mean my site has authority?
Microsoft says no. It states the metrics reflect “how often your content is referenced” without indicating “ranking, authority, or the role of any page within an individual answer”.
What are grounding queries?
Microsoft describes them as the “key phrases the AI used when retrieving content” — the system's own internal search terms, not necessarily what the user typed.
Do the operators explain how sources are selected?
No. Google publishes eligibility conditions, OpenAI publishes crawler controls, and Microsoft publishes outcome data. None publishes selection criteria.
Will a placement on another site show in my AI Performance report?
No. The report covers sites you verify in Bing Webmaster Tools, so citations of a publisher's page accrue to that publisher, not to you.
Can anyone optimise for AI citation?
Only at the level of preconditions — reachable, indexed content that answers specific questions. No published criteria exist to calibrate anything more precise against.
Sources
- official Microsoft. Introducing AI Performance in Bing Webmaster Tools, public preview announcement, for citation metrics, grounding queries and coverage.
- official Google Search Central. Documentation on AI features in Google Search, for eligibility conditions for AI Overviews and AI Mode.
- official OpenAI. Crawler documentation for GPTBot, OAI-SearchBot and ChatGPT-User.