GEO measurement: distinguish brand mentions, citations and referral visits
Build a reproducible observation log. Separate appearance, links, recommendations and website visits, and keep technical audit scores out of citation probability claims.
Define the event before choosing a metric
No single GEO number represents every outcome. A brand appearing in an answer, a link to a particular page, a recommendation to buy and a visit to the website are different events. A technical audit inspects whether public pages are accessible and understandable. Combining these observations into one “AI ranking” conceals useful detail.
| Event | Minimum evidence | What it does not establish |
|---|---|---|
| Brand mention | The full answer names the brand in context | A recommendation or link to the official site |
| Page citation | A usable citation URL with the final page checked | A click or purchase |
| Product recommendation | The actual advice and its conditions | Consistent recommendation for every question |
| Referral visit | A recorded visit, source or explicitly labelled test click | Complete attribution when referrer data is missing |
| Technical readiness | HTTP, robots and page-content observations with a date | Indexing or a particular citation probability |
Establish a question baseline you can compare
Use questions drawn from real decisions: who the product suits, when to choose it, what to compare and where its limits lie. Save the question set, language, region, model, browsing setting and whether you supplied candidate URLs. Preserve these conditions in later observations and record necessary changes. Do not combine an unprompted brand appearance with a response that followed an explicit request to inspect the brand’s URL.
Save the complete prompt and answer, timestamp, model, cited URLs and interpretation. Separate failed requests, timeouts and incomplete answers from valid answers with no citation. State sample size and repetition count. “Three of twelve valid answers cited the target page” describes that observation set; it is not a market-wide recommendation rate.
Check what the linked page actually supports
Open the cited URL and inspect the final destination, page subject and claim supported. The same brand name may identify another business. A link to an independent review is not an official-site citation. Label observations as official-site citation, external-source mention, brand-only appearance or uncertain, and retain the original answer.
A recommendation can also be conditional. “Useful for small batches, but consider a cheaper manual option first” differs from “the first product to choose”. Preserve qualifications instead of counting every appearance as a positive recommendation.
Keep website visits in a separate record
In the analytics system, retain the source, landing page, observation window and available campaign parameters. Label your own verification clicks as tests. A successful test shows that collection works, not that customers arrived. Copying a URL, privacy settings and opening links from an app can remove referrer information. Direct traffic cannot all be attributed to AI, or all ruled out as AI traffic.
The existing AI referral traffic guide explains practical collection and testing. Separate real referrals, manual tests and unknown sources. Do not turn a simulated source header into a growth claim, or assume that an AI tool opening a page represents a customer visit.
Record the change before interpreting an effect
For each website update, record the canonical URL, question addressed, evidence added, modification time and content version. In a before-and-after comparison, explain other changes such as model versions, the question set, region or traffic mix. A citation occurring after an edit does not by itself establish that the edit caused it.
Useful case studies include unsuccessful results. Show the method, underlying evidence, observation window, sample size, limits and next decision. Use clearly labelled illustrative examples when customer permission or complete records are absent. Do not invent customer identities, traffic lifts or conversion numbers to fill a case-study template.
Make a content decision from the evidence
Fix the relevant page when there is an access problem. Improve the answer and its evidence when an independently useful question is poorly covered. Investigate source attribution and the reader’s reason to click when citations do not produce visits. When existing pages already cover the intent and proposed drafts are repetitive, research another question before writing more. Measurement should help choose the next useful action, rather than manufacture a flattering combined score.
Sources and further reading
- Google Search Central — AI features and your website
- OpenAI — Crawlers and user agents
- Google Analytics Help — Traffic acquisition report
The workflow examples are editorial illustrations, not customer results or verified performance claims. Platform documentation describes platform behaviour; it does not endorse this guide.