Gemini, Microsoft Copilot and Yandex: the surface-specific checks

Three surfaces that expose very different evidence

Gemini, Microsoft Copilot and Yandex are often discussed as though they were interchangeable competitors to ChatGPT. For an audit, the differences matter more than the similarities. What can be known about each product, and from where, is different, and an honest report has to say so.

Gemini: 21 specific checks

Gemini is available as a consumer app and as a developer interface. They are different surfaces with different documented behavior, and a result from one cannot be credited to the other. GEMINI-37 states the match limit between the two, and GEMINI-23 to GEMINI-25 and GEMINI-28 to GEMINI-31 cover the model interface in detail: whether a search tool was requested, whether a search event really happened, which queries were documented, how text segments tie to sources, and whether the response was parsed with a parser that fits the true structure of the output.

Grounding and citation

Gemini can ground an answer in Google Search. The consequences for an audit are practical. A source that appears as a related link is not necessarily a source that supports a claim, and GEMINI-33 keeps those two apart. GEMINI-30 checks that each citation attaches to the right text block in valid order, and GEMINI-34 stores search suggestions and citation presentation in their raw form so that nothing is reconstructed from memory. GEMINI-22 handles the Google verification badge separately from ordinary source citation, because the badge is a different kind of evidence.

Which part of the app is the source?

The Gemini app shows a source panel, inline links and related links. GEMINI-17 parses them into types, and GEMINI-32 separates related links, Google Search, app redirects and supporting sources. Without this separation, a report could count a navigation link as an endorsement.

Preferences that look similar but are not

GEMINI-20 records your Google-Extended preference and what it does and does not affect. GEMINI-36 treats that preference and Google Search’s own AI control as two separate policies. A site owner can hold a firm view on one and a different view on the other. Merging them into a single “block Google AI” decision hides a real choice, so the report keeps them apart.

Accuracy of facts about people, prices and brands

GEMINI-18 checks the person and organization relationships that an answer states against dated, correct information. GEMINI-19 traces an old price to the source passage it likely came from, and asks whether package, currency or period explains it. GEMINI-27 asks whether a wrong price or fact originates in an inconsistency between language versions or in an outdated date. GEMINI-26 verifies that a brand’s profiles and alternative names point to a single entity. These four records are the most directly useful for businesses that have been misdescribed.

Country, queries and site-directed research

GEMINI-21 records the requested, transmitted and independently verified country as three separate facts, because a country setting that was requested is not proof that it was applied. GEMINI-35 keeps Gemini’s natural discovery apart from research that has been pointed at your site through a URL context feature. The second is a test of what the model can read on your site when directed. It is not evidence of how it behaves on its own.

Microsoft Copilot: 23 specific checks

Copilot is closely tied to Bing. That has two consequences. First, Bing’s crawler, index and webmaster tools are the main levers for a site owner. Second, Bing offers owners something many other surfaces do not: an AI performance view in Bing Webmaster Tools. CiteMoment uses it extensively, with care about what it can and cannot show.

Access and indexing

COPILOT-17 separates robots permission for Bingbot from real access through your firewall. COPILOT-23 records the page’s status in Bing’s index through a documented route or the owner’s export. COPILOT-24 and COPILOT-39 cover IndexNow: whether changed addresses were reported to participating search engines, whether the key and domain ownership are valid, and what response was accepted. IndexNow is a notification mechanism, not a guarantee of indexing, and the report treats it as exactly that.

NOARCHIVE and NOCACHE

COPILOT-18 to COPILOT-20 and COPILOT-38 read the NOARCHIVE and NOCACHE directives. They are recorded separately, their combined behavior is labeled with its documented scope, and Bing’s current behavior is checked against dated documentation rather than assumed. COPILOT-38 reports separate scope for Google, classic Bing and Copilot, since a directive that means one thing in one product may not mean the same in another.

Bing Webmaster Tools: use it, but read the fine print

COPILOT-25 to COPILOT-29 and COPILOT-33 to COPILOT-37 are about the AI performance data. They retrieve total citations, cited pages, the daily page count, a sample of the queries Copilot grounded on, intent and topic classifications, and citation share with a period comparison. Each comes with a limitation that the catalog insists on stating.

• The total may combine several surfaces (COPILOT-33).
• The average of cited pages must be matched to the official daily definition (COPILOT-34).
• Grounding queries are a limited sample, not all queries (COPILOT-35).
• Citation share compares your citations with those of all sites and says little about a specific competitor (COPILOT-36).
• Intent and topic classes are suggestions with a confidence level, not facts (COPILOT-37).

Surfaces, sources and success

COPILOT-30 keeps Copilot chat and Copilot Search in Bing as separate series. COPILOT-31 keeps enterprise, consumer and private data sources apart from open web results. COPILOT-32 distinguishes three states for a web source: cited, merely listed or unmarked. COPILOT-22 counts delivery failures, ambiguous citations and genuine absence separately so that the denominator of “successful measurements” stays honest. COPILOT-21 traces an outdated address back to a source passage and compares it with the business’s correct address.

Yandex (Russia, yandex.ru): 22 specific checks

Yandex is the most specific surface in the catalog because it is bound to a market. CiteMoment’s Yandex records concern the Russian search service at yandex.ru. YANDEX-19 states plainly that Yandex’s consumer product in Turkey cannot be assumed to stand in for the search service in Russia, and YANDEX-27 asks whether the selected market and the supported measurement path agree with each other. A measurement made from the wrong market is not a measurement of Yandex Russia.

Crawlers and robots

YANDEX-20 shows which quick-answer product is affected by your permission for the YandexAdditional and YandexAdditionalBot robots. YANDEX-21 records which address YandexBot really requested, when, and with what response code. YANDEX-37 states which Yandex product your robots preference affects and the scope of its documented update. YANDEX-38 combines evidence of access permission, crawling, indexing and AI attribution for the same address and time, because each one alone leaves a gap.

Webmaster data and Alice AI

YANDEX-22 imports index data from Yandex Webmaster with its source and timestamp. YANDEX-23 reports Yandex’s official Alice AI visibility share with its valid query scope and period. YANDEX-24 treats the query and source examples Yandex shares as a limited sample, not as all queries, and YANDEX-35 explains how the official share differs from a rate computed from a private query sample. YANDEX-25 and YANDEX-36 classify the neighboring sites that appear for near queries as guides, media, official sources or competitors, rather than assuming they are all rivals.

Products, languages and warnings

YANDEX-29 records whether a result came from Alice, search, quick answers, GenSearch or a Turkey product. YANDEX-30 verifies from live documentation which language Yandex actually produces answers in, and YANDEX-28 separates a missing site language as a possible influence from a proven reason for not appearing. YANDEX-26 and YANDEX-34 keep three warning classes apart: source not found, inference failure and quality suspicion. YANDEX-32 and YANDEX-33 deal with interim and final responses and the rich and plain response schemas. YANDEX-17, YANDEX-18 and YANDEX-31 record search type, query language, region, domain and address filters.

The shared contract behind all three

The 30 shared records, described in the catalog explanation, apply to each surface here exactly as they apply elsewhere. Because Gemini, Copilot and Yandex each disclose less than their owners might wish, the shared rules about unknown conditions are especially important. COMMON-03 separates known from undisclosed device and chat conditions. COMMON-10 marks hidden internal searches as unknown. COMMON-M22 distinguishes refusals, timeouts and empty answers. COMMON-M23 and COMMON-M24 insist on denominators and uncertainty.

The result is a report that is sometimes less tidy than a competing tool’s dashboard. It contains more “partial” and “unmeasurable” labels, because it refuses to fill gaps with plausible numbers.

Practical steps for each surface

For Gemini

1. Decide your Google-Extended and Search preferences deliberately and separately.
2. Make prices dated and consistent across language versions.
3. Keep brand profiles and alternative names consistent and linked with accurate sameAs references.

For Microsoft Copilot

1. Verify your site in Bing Webmaster Tools and review the AI performance reports, noting their limits.
2. Allow Bingbot, check that your firewall lets it through, and implement IndexNow for changed pages.
3. Review NOARCHIVE and NOCACHE usage so you know what each one does in each product.

For Yandex

1. Decide first whether the Russian market is part of your business. If it is not, treat this surface as informational.
2. If it is, verify the site in Yandex Webmaster and review robots rules for Yandex crawlers.
3. Confirm that your language coverage matches what Yandex supports, and avoid assuming that a missing language is the proven reason for invisibility.

For general principles that apply to all six surfaces, read the AI search visibility guide. For the other three surfaces, see ChatGPT Search, AI Overviews and AI Mode. To have these checks run on your site, contact Smartise.

Frequently asked questions about Gemini, Copilot and Yandex checks

What is the difference between Google-Extended and Google Search controls?

They are two separate policies. Google-Extended is a preference concerning Gemini training and grounding, while Google Search has its own controls for how pages appear in search and AI features. CiteMoment reports them as separate records, GEMINI-20 and GEMINI-36.

Why does CiteMoment separate the Gemini app from the Gemini API?

They are different surfaces with different behavior and documentation. Results from one cannot be credited to the other, so CiteMoment compares them as separate series with their limits explained.

What does Bing Webmaster Tools tell me about Copilot?

It provides AI performance data such as total citations, cited pages, a sample of grounding queries and citation share. Each metric has limits, for example that totals may span several surfaces and that grounding queries are a sample.

Does IndexNow guarantee that Bing indexes my pages?

No. IndexNow notifies participating search engines that addresses have changed. CiteMoment checks the key, the domain ownership and the accepted response, but treats it as a notification, not a guarantee of indexing.

Is Yandex relevant if my customers are outside Russia?

Probably only slightly. CiteMoment’s Yandex records concern the Russian search service at yandex.ru, and the Turkey consumer product cannot be assumed to represent it. If Russia is not a market for you, the surface is informational.

Why are some Copilot, Gemini or Yandex items reported as unmeasurable?

Because the provider does not disclose the condition or the data path is not available. Unmeasurable items are listed with their reason instead of being replaced by estimates.