AI search visibility · Playbook
AI-search guidance — what Google actually says, and what that means for a AI search visibility report
Issue #71. chatot measures visibility well — is the business named, cited, list-placed. What it did not do was tie its recommendations to first-party guidance, which matters more here than in the rest of the lane, because “GEO” and “AI SEO” attract more folklore per square inch than any other corner of search.
Primary source: Google — AI features and your website · last verified 2026-08-06.
Start with the sentence most GEO advice omits
There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.
Google’s stated position is that eligibility for AI features is eligibility for ordinary Search snippets. No AI-specific markup, no new machine-readable file, no schema type that unlocks it.
This is the most useful thing in a chatot report, and it is a commercial advantage rather than a disappointment. A client who has been pitched an “AI optimization package” by someone else can be told, with a citable source, that the lever does not exist — and that the work that does move AI visibility is the work already on their plan. A recommendation implying a special AI lever is unsupported by the vendor whose product it targets, and it will not survive contact with a client who checks.
What Google does say to do
All of it is ordinary, and all of it is already covered by skills in this lane:
| Guidance | Covered by |
|---|---|
| Page must be indexed and snippet-eligible | orbeetle (noindex, canonical), ariados (robots.txt) |
| robots.txt allows crawling | ariados |
| Content is discoverable through internal linking | galvantula |
| Important content is in text, not images or video alone | oranguru (thin content) |
| Strong page experience | rapidash, conkeldurr, axew |
| Structured data matches the visible text | porygon |
| Business Profile and Merchant Center kept current | nosepass (NAP, GBP readiness) |
That mapping is the honest answer to “what do we do about AI search?” — and it is why chatot composes rather than duplicates.
Controlling what appears
Existing snippet controls, not new ones:
nosnippet,data-nosnippet,max-snippet— limit what can be shown.noindex— removes the page from Search, and therefore from AI features.
Worth stating to a client considering these: the controls are all-or-nothing against Search itself. There is no directive that keeps a page in the ten blue links but out of an AI Overview, so using them to opt out of AI features costs the ordinary ranking too.
Measurement — and chatot’s honest limits
Traffic from AI features appears in Search Console’s Performance report under the Web search type. It is not broken out separately, so a report that claims to isolate “AI traffic” from Search Console is claiming a number Google does not publish.
What chatot measures is different and should be described as what it is: live sampling of engine answers to the business’s own prospect questions. That is a genuine, reproducible signal about whether the business is named, cited and list-placed — and it is a sample of answers, not a traffic figure. Answers vary by phrasing, by user context and over time, which is why the report tracks change across runs rather than presenting one run as a score.
The rule for writing recommendations in config.json
Each recommendation should carry source and lastVerified when it rests on published guidance, the
same contract every finding in the lane uses since 0.102.0. Concretely:
- Ground it. If the recommendation maps to a row in the table above, cite this page.
- If it does not map, say so. Some good advice is inference rather than guidance — “get cited by the directories the engines actually quote” is a reasonable strategy drawn from chatot’s own cited-source data, not from Google. Mark it as observed, and point at the run’s own evidence.
- Never imply a lever Google says does not exist. No “AI schema”, no “LLM meta tag”, no “optimize for the AI crawler” beyond ordinary crawlability.
Non-Google engines (ChatGPT, Perplexity) publish far less, and what they publish changes quickly. Where a recommendation is engine-specific, cite that engine’s own documentation and date it — an undated claim about an engine’s behaviour ages badly within months.