Key takeaways
- Measure citations, referrals, and platform reports — not a synthetic GEO score nobody can audit.
- A fixed query panel plus a citation log beats dashboard vanity. Disclose volatility when models change.
- Search Console generative-AI reporting is evidence when you have it. It is not a substitute for useful pages.
A vendor emailed a GEO score of 74. They would not share the query set, the geography, or the model version. The homepage still says “innovative solutions.”
Answer-engine visibility is measurable without inventing a number. Use a query panel you choose, a citation and referral log you maintain, and platform evidence such as Google Search Console generative-AI performance reports when your property has access. Volatility is part of the method. A made-up GEO score is not measurement. It is a product selling certainty the channel does not offer.
What generative engine optimization actually is is the definition: truthful pages easy to quote. SEO and GEO for small businesses is the alignment rule. This page is how you track whether that alignment is showing up in generated answers — without minting authority in a spreadsheet.
Visibility is citation, not a badge
An answer engine does not rank you on a leaderboard it publishes. It retrieves or repeats short claims from sources it trusts for a specific question at a specific time. Measurement therefore looks like:
- Did a named query produce an answer that mentions the business correctly?
- Did the answer cite or link to a URL you control?
- Did qualified traffic or branded search move on a lag you can explain?
That is different from classic SEO position tracking. Positions for ten keywords miss the point when the interface is a paragraph, not ten blue links. It is also different from social listening. A mention in a thread is not a citation in an answer card.
| Signal | What it tells you | What it cannot tell you |
|---|---|---|
| Citation log | Whether you appeared for chosen queries | Total market share |
| Referral traffic | Whether users followed an AI surface to you | Revenue causality alone |
| Search Console AI reporting | Aggregated impressions/clicks from Google’s AI experiences | Competitor detail |
| Branded search lift | Whether more people look for you after exposure | Which sentence was quoted |
Google’s resource on optimising for generative AI in Search repeats the foundation: helpful, people-first content and ordinary SEO health still matter. Measurement should test whether that foundation is quotable — not whether a vendor’s opaque index moved three points.
Build a fixed query panel
Choose twenty to forty queries that reflect real buyer jobs — not only your brand name. Mix:
- Category: what you do, in language a stranger would use.
- Comparison: you vs alternatives you actually compete with.
- Local or niche: territory, method, or constraint you publish honestly.
- Branded: misspellings and “is X legit” patterns.
Run the panel on a schedule: weekly for active programmes, monthly for stable sites. Record date, interface (Search AI overview, chat product, other), exact query, whether you were mentioned, whether a URL was cited, the quoted text if any, and screenshot or export reference.
Keep the panel fixed for at least a quarter before you swap queries. Changing the panel every week is how teams manufacture improvement.
Thursday: run five non-branded queries you care about. Log mention yes/no, citation yes/no, and the sentence quoted. If the answer invents facts about you, fix the page — not the log.
Maintain a citation and referral log
Parallel to the panel, track inbound evidence:
- Referrers from AI products in analytics — filter new referrers monthly.
- Links from third-party answers when you find them in the wild.
- Support and sales notes when prospects say “I saw you in ChatGPT” — tag them; qualitative, but useful.
Each row should include date, source, URL cited, claim repeated, and whether the claim matches your public copy. Mismatches are content bugs. Invented praise is worse than silence — it means the model guessed.
Disclose volatility in every report you write internally. Model versions, geography, and signed-in state change outputs. A citation on Tuesday may be absent on Friday without any change on your site. That is not failure of the log. It is the channel.
Avoid synthetic “AI visibility scores” that hide methodology. If a tool cannot show query set, sampling rate, locale, model family, and refresh cadence, treat it as directional entertainment — not a KPI.
Use Search Console when you have it
Google has begun rolling out generative-AI performance reporting in Search Console for properties that receive impressions in AI experiences. When available, use it as platform-grounded evidence: impressions, clicks, and queries that surfaced your URLs in those contexts.
Read it beside ordinary Search performance, not instead of it. AI impressions are not revenue. They are a signal that a URL was eligible to appear. The page still has to do the job when the click arrives.
If your property does not yet have the report, you are not behind a paywall — you are early. Rely on the query panel and referral log until platform data exists. Do not backfill the gap with invented scores.
What good measurement changes
Measurement should change copy and structure, not slide decks.
- Wrong entity quoted: align title, heading, body, and schema on the same facts.
- Right entity, wrong offer: rewrite the first screen as sentences a cautious model can lift.
- Cited URL is not the money page: fix internal links and canonicals so the useful page is the one retrieved.
- Mentions without clicks: test whether the cited page has a next step worth taking.
SEO content is how those fixes ship as durable pages. SEO is the ongoing programme when the panel and logs are part of quarterly review, not a one-time panic after a vendor email.
The estimator at get started scopes the work from evidence you can show — citations logged, volatility noted, pages worth quoting.
