How to measure AEO: mentions, links and enquiries
Neaptide · September 7, 2026 · 6 min read
Choose AEO/GEO metrics, account for failed checks and compare observations fairly. A synthetic dataset makes the example calculations reproducible.
On this page
“Visibility reached 40%” is an incomplete report. Which questions were tested, what counted as success and how many answers were collected? Without those details, the percentage might change because difficult questions disappeared from the sample.
Measure brand presence, domain links, factual accuracy and visitor actions separately. The first three can be studied against a fixed question set. Visits and enquiries require their own analytics; they cannot be calculated from mention counts.
Fix the question set before starting
Build the question set from customer conversations, searches on your site, support requests and available search reports. Group questions by purpose: understanding a topic, comparing options or choosing a supplier. Track questions that name your brand separately: an answer to a question about your company is different from a recommendation in response to a general enquiry.
This is your research panel, not every question in the market. Without demand volumes, do not call its results market share. Preserve the original list and add new questions as a separate group rather than rewriting the basis of past comparisons.
Record the context of every observation
- Exact question, language, target market and whether it names the brand.
- System, available model name, search mode, date and repeat number; mark unknown settings.
- Full answer and actual source URLs. Screenshots help with context but do not replace links.
- Status: usable answer, absent answer feature or technical failure. Keep these distinct.
- Separate labels for brand mention, own-domain link and factual accuracy, using agreed rules.
A worked example: choose the denominator
These are synthetic teaching data, not requests run against real systems or a client website. Assume 20 questions tested three times: 60 attempts. Six fail technically and 54 return usable answers.
| Metric | Calculation | Meaning |
|---|---|---|
| Technical failures | 6 / 60 = 10% | Attempts without a usable response |
| Brand mentions | 18 / 54 ≈ 33.3% | Collected answers containing the brand |
| Own-domain links | 9 / 54 ≈ 16.7% | Collected answers linking to the domain |
| Accurate facts among mentions | 16 / 18 ≈ 88.9% | Two mentioning answers contain a factual error |
A tool failure does not establish brand absence, so failures are shown separately. Dividing 18 mentions by all 60 attempts gives 30%, a different metric. Comparing 33.3% with last month while hiding failures would also mislead.
Check how many answers were collected for each question. A question with three responses contributes more to the overall percentage than one with only two. If failures are concentrated in an important group, collect the missing responses or compare only observations available in both periods. All 60 synthetic records are available to download and recalculate.
Compare before and after carefully
Keep wording, languages, modes and repeat schedules consistent. Record deployment dates separately. A changed model or collection method needs a marked break and an explanation. Repeating a question reveals variation; it does not automatically produce independent statistical observations.
A higher percentage alone does not establish that your changes caused it. System behaviour, competitors and sources can also change. Report the observation separately from causal hypotheses, then inspect the answers: did the desired link appear, or merely the company name?
Track visits and enquiries separately
Use available search-platform and website analytics reports with their own metric definitions. Google documents measurement for its AI features in Search Central; those data do not represent other systems.
Separate visits with an identifiable source from conversion events. An unattributed visit is not automatically AI traffic. A customer saying they found you through AI is useful self-reporting, but differs from a technically recorded referral.
Make the report actionable
Put completed changes, observations and next actions together. For example: the delivery timeline was corrected, some answers still show the old timeline, so inspect their sources. This defines work more clearly than “increase our AEO score”.
With Neaptide, agree the measurement scope alongside promotion goals: priority services, questions, data access and owners for fixes. A small repeatable panel with clear findings can be more useful than an impressive percentage nobody can reproduce.