An AI answer can shape the shortlist before the user ever visits the site. Wrong price, availability or delivery data can remove the store from consideration before the first click.
Pilot ready
GEO: measure what AI systems see before and after changes
In an AI answer, simply being mentioned is not enough. A store can appear without a citation, point to the wrong page, or have price and availability read incorrectly. I therefore measure four separate things: presence, source, fact extractability and fact accuracy.
Case at a glanceDesigned the measurement method, control-query set and page requirements so the result can be compared before and after changes.
There is a repeatable 60-query measurement design with Day 14 and Day 30 checkpoints. The first extraction test already showed different behavior across AI systems.
01Baseline→02Changes→03Repeat→04Compare
The same query set is repeated after changes, so the result is a time comparison rather than one lucky answer.
A store can be shortlisted before the click
An AI answer can surface price, availability, delivery and a source before the user ever reaches the site. Brand presence alone is therefore not enough: presence, fact extractability, source and accuracy need separate measurement. Otherwise a visible brand mention can still carry wrong commercial facts.
Key finding03
The correct conclusion is to test extractability per client and scenario.
One test showed different results across AI systems
In the same scenario, different AI systems retrieved different commercial facts from the site. That made the problem concrete: availability has to be tested, not assumed.
| Check | Expected | Observed |
|---|---|---|
| GPT / Claude / Googlebot-Extended | Retrieve commercial facts from the page | The test returned two products and exact prices. |
| Perplexity | Retrieve the same commercial facts | In this test the system reported that it could not render the site. |
Why a brand mention is not enough
An AI answer can name a retailer without citing it, or cite a page while reading the price or availability incorrectly. Presence, fact extraction, source citation and accuracy are therefore measured separately.
What is actually measured
Each metric answers a different question.
Presence
How often the brand or store appears across the agreed query set.
Fact availability
Can the system retrieve price, availability, delivery and other important conditions?
Source used
Which pages and snippets are used as the basis of the answer?
Accuracy
Do price, availability, delivery and other conditions match the site?
What needs to be easy to retrieve
We check separately whether a fact can be retrieved and whether it still makes sense outside the full page.
Important facts in HTML
Price, availability, delivery, the buy action and navigation are available in the source page.
Short, self-contained answers
A key fragment keeps its meaning outside the full article: answer, condition and source.
Consistent relationships
The company, product, offer, article and policy information point to the same underlying entities.
How changes are measured
First establish the baseline, then apply changes and repeat the same sample.
Baseline: 60 queries
12 queries × 5 categories; record presence, cited pages and fact accuracy.
Improve the pages
Fix the issues found in fact availability, content and entity relationships.
After 14 days
Repeat the same scenarios and check the first movement.
After 30 days
Repeat again and decide which changes are worth scaling.
How the 30-day pilot is structured
The program repeats the same scenarios before and after changes so that a lucky one-off answer is not mistaken for progress.
60 queries
Twelve user scenarios across five categories create the baseline.
Day 14
The first repeated check after technical and content changes.
Day 30
The second checkpoint and the decision on whether there is enough evidence for a 90-day scale-up.
Four KPI families
Share of Voice, extractability, cited URL/snippet, and accuracy of price/availability/delivery are measured separately.
Context & measurement
- Context
- GEO / AI Search program for e-commerce
- Scale / period
- 60-query baseline · Day 14 / Day 30 · validation pending
Next step
Run the baseline, ship the pilot changes and compare the Day 14 and Day 30 checkpoints.
AI Search / GEO
Facing a similar challenge?
Tell me where SEO is blocked by product, engineering or data.