Seven AI surfaces: what public evidence can actually verify

We ask what an outside observer can establish without access to indexes, ranking signals, or model context. This is desk research over official documentation, not an experiment with matched prompts. Conclusions therefore describe public capabilities and limits, not visibility shares.

Seven surfaces may look like one AI question to a user, but they are different research environments with different documents and signals. We do not rank them: the map shows what an operator describes, what can be observed, and what remains unknown.

For readers, the map is a decision aid: first identify whether the question concerns access, search, attribution, or referral, then choose the measurement. This prevents transferring one operator’s rules to another.

Check attribution quality separately: a link may lead to a homepage, a secondary summary, or the exact passage. These are three different outcomes. Record the URL beside the supported claim so any domain appearance is not counted as a citation.

Platform comparison requires the same unit of analysis. We use a surface and a concrete observable artifact, not a brand as a whole. YandexBot, OAI-SearchBot, and a search interface must not be combined without an explicit definition.

The editorial value of the map is choosing the next action. A bad HTTP status leads to a technical check; a missing index signal to an indexability check; a relevant answer without a link to a separate attribution experiment. One universal tip would be less precise.

Show the desk-research status beside the date: this is a document synthesis, not answer measurement. Readers can then see why the table maps knowledge boundaries rather than traffic or ranking shares.

Section sources:[1] Google Search Central

The table deliberately separates public description, observation, and unknowns. A row is an evidence boundary, not a platform quality score.

Read the table by row and column: choose a surface, separate operator description from observation, then check the final column. A URL in an answer can be checked; the full candidate set before selection is not public.

When a document discusses a crawler, we do not add a claim about the answer engine. When an interface shows a source, we do not declare the whole retrieval proven. This discipline makes comparison less flashy but audit-ready.

For comparison, identify the artifact type: server line, console record, full answer, or referral. Artifacts of different kinds cannot be treated as interchangeable evidence.

Section sources:[2] OpenAI Developers[1] Google Search Central[3] Perplexity Docs[4] Microsoft Learn[5] Google Gemini Help[6] Anthropic Support[7] Yandex Webmaster

For each surface we record the official URL, publisher, review date, and wording. We code four fields: discovery, fetch, attribution, and unknown. Industry advice without a supporting source is excluded from observed claims.

Synthesis does not mean seven documents describe one pipeline. Preserve the canonical URL, review date, and exact claim wording, then label observed, inference, or unknown. An editorial recommendation is not an operator fact.

A useful record might say: “September 11, RU, Google AI Search, query X, URL Y displayed, source partially supports claim Z.” It is stronger than “the page is visible in AI” because it names surface, date, link, and match strength.

Check every source date and scope. An outdated access policy may explain a log mismatch; create a new claim version instead of rewriting the old observation.

Section sources:[3] Perplexity Docs

The portable recommendation is not a special AI file but accessible HTML, stable URLs, clear entities, primary sources, and separate outcome measurement. This reduces verification friction but does not guarantee display or citation.

In practice, fix delivery first: HTML, stable URLs, sitemap, and response codes. Then create evidence-bearing material with dates, sources, and limits. Only then measure a surface answer. Citation absence cannot be assigned one cause.

The next study should compare matched questions rather than collect random screenshots. It needs preregistered citation criteria, preserved answers, and a document-change log. Only such a slice can support a discussion of repeatability.

The editorial action after the map is to choose one testable question and one URL. A broad list of tips creates an illusion of coverage, while a narrow repeatable test yields stronger observation.

Section sources:[4] Microsoft Learn

Documentation may cover only part of a product and may change. We did not run matched prompts, measure ranking, or test causality between crawling and answers. Any quantitative conclusion requires a separate protocol and preserved answers.

A public feature description is not an observation in every language, region, and mode. We therefore publish no numbers without denominator, period, and sample definition. This sets a standard for future experiments.

Implementation is simple: store sourceIds, data-through date, claim type, and correction history for every publication. The public page then shows not only the conclusion but its verification path.

Publish an absolute run count beside the rate and list excluded answers. Without a denominator, readers cannot assess result stability.

Section sources:[1] Google Search Central

Repeat the review in 30 days and preserve document snapshots and hashes. Then ask ten matched questions on each surface, run three repetitions, and record mode, region, answer, and links. Count fetch, mention, citation, and referral separately; record missing artifacts as unknown.

Protocol: preregister questions, choose the same locale and region, save the full answer, open URLs, check semantic match, and export server events. Count fetch, index signal, mention, citation, and referral separately.

We deliberately do not call an unknown a platform failure. The absence of a public candidate set is a property of available evidence, not a measured error. The study helps formulate the next experiment correctly.

A repeat run must preserve conditions: prompt, language, region, mode, date, and interface. Otherwise context change will be mistaken for visibility change.

Section sources:[1] Google Search Central