AI search this week: sources, crawlability, and measurement

Dedicated reporting for Google’s generative features reduces reliance on guesswork but requires disciplined interpretation. A Search Console impression belongs to the measured Google surface. It cannot be transferred automatically to ChatGPT Search, Perplexity, or Claude, and it is not proof of a particular citation without an archived answer.

Reading the two documents together gives a practical boundary: Google describes measurement inside Search, not a universal AI-visibility counter. A weekly report should therefore keep separate series for Search Console, an archived answer panel, and server logs. A shared trend is a reason to investigate, not a combined metric.

Section sources:[2] Google Search Central Blog[1] Google Search Central

OpenAI and Perplexity documentation shows why analytics needs purpose-based classification. A search crawler, a crawler used for potential training, and a user-triggered fetch create different events. One “AI bots” chart may look persuasive but cannot explain why a page was requested or whether it was shown to a user.

The practical result is not to allow or block “all AI bots” with one rule. Store role, IP verification, and agent purpose separately. Even a verified request means only that access to the resource was observed; it does not reveal what the system extracted or how it answered.

Section sources:[4] OpenAI Developers[3] Perplexity Docs

Google continues to recommend Search fundamentals: crawl access, indexable text, internal links, and useful content. This is not a promise of inclusion in an AI answer. It is a set of conditions whose absence creates a clear obstacle before any private retrieval stage.

It is therefore useful to separate the work into two layers. First check robots.txt, HTTP status, canonical, indexable HTML, and internal linking. Then observe platform answers separately. Fixing the first layer improves technical eligibility but cannot promise citation in the second.

Section sources:[1] Google Search Central

Take ten important URLs and create six columns: HTML available, canonical correct, index status, verified fetches, archived mentions, and archived citations. Do not replace missing observation with zero; use unknown. The table quickly distinguishes a technical problem, a measurement gap, and a genuine absence from answers.

A repeatable check is one short cycle: freeze ten URLs and ten questions, export Search Console for an equal window, check HTML, archive platform answers on one day, then repeat a week later. Record the observation source and confidence in every row. A missing answer must not be turned into zero visibility.

For the next cycle, choose ten URLs and ten questions. Record HTML, canonical, robots.txt, and sitemap state; export an equal Search Console window; archive platform answers with date, language, and mode; and match server requests by verified agent. Keep observation, source, limitation, and next test as separate columns. This lets the editorial team distinguish missing data from a zero result.

Section sources:[2] Google Search Central Blog[1] Google Search Central[4] OpenAI Developers[3] Perplexity Docs

Before publishing a conclusion, answer five questions: which source supports the claim; which surface does the metric belong to; do period and language match; is there an archived answer or only a log; what remains unknown after the check? If any answer is unclear, label the result an observation or hypothesis and schedule a recheck.

Before release, check five points: the fact has a primary source; the source belongs to the same surface; period and language match; the conclusion is no stronger than the data; and unknown is not replaced with zero. If an interface or document changes, start a new observation series and preserve the old one as a separate version. The digest connects signals without erasing their differences.

Section sources:[2] Google Search Central Blog[1] Google Search Central[4] OpenAI Developers[3] Perplexity Docs

Verify report availability in real accounts, fetch official IP endpoints again, and preserve document hashes. For any prompt panel, freeze the question set, modes, languages, and repeat count. If a platform interface changes, begin a new series rather than merging incomparable answers.

In the next issue, the team will repeat the same URL and question set after checking the documentation. Results will be recorded as separate observations for crawling, reporting, and answers. This decision reduces the risk of treating source access as proof of source use.

Three signals form one practical picture. Google’s reporting adds observability inside Search, OpenAI and Perplexity documentation requires separating request purposes, and Search fundamentals remain a technical entry condition. Together they are not one visibility score: they are three layers—surface measurement, agent access, and answer observation. A weekly conclusion should first state where the signal changed and only then propose a check.

If only crawler requests rise, that indicates a change in access or agent interest, not higher citation. If Search Console impressions change, that belongs to the Google surface. If a link appears in an archived answer, that is a separate observation. The digest aligns these events in time but does not add them into one score.

This week’s editorial decision is to publish a measurement protocol, not an “AI optimization” promise. New materials add an evidence-boundary block, section-level source IDs, and explicit unknown labels. The editorial queue accepts claims only when they can be tied to official documentation or a reproducible archive. Practitioner opinions remain hypothesis signals and do not replace facts.