RESEARCH METHOD · METRICS
AI search metrics and statistics you can trust
AI search numbers often mix platform facts, vendor panels, one-time prompt tests, and forecasts. Use a source and denominator checklist before a statistic influences product, content, or budget decisions.
The short answer
A trustworthy AI search statistic states who produced it, what population was measured, when it was collected, which engines and locales were included, how prompts were sampled, how failures were treated, and what the number does not prove. If those fields are missing, treat the number as directional rather than factual for your market.
1. Classify the source
| Evidence class | Useful for | Main limitation |
|---|---|---|
| Official platform documentation | Eligibility, controls, product behavior, metric definitions. | Usually does not quantify your category. |
| Peer-reviewed or disclosed primary research | General patterns under a stated design. | May not transfer to current engines or your market. |
| Vendor study | Large operational samples and trend hypotheses. | Sampling and commercial incentives need review. |
| Your controlled observations | Decisions for your brand, prompts, locale, and time window. | Often small and not a market-share estimate. |
| Forecast or commentary | Scenario planning. | Not observed behavior. |
2. Require a denominator
“Citation rate was 40%” is incomplete. A useful statement is “the brand received a visible citation in 8 of 20 valid answers across a frozen prompt set, with 3 additional platform failures reported separately.” The denominator, failure handling, engine mix, locale, and date determine what the percentage means.
3. Do not combine incompatible metrics
Search Console impressions, analytics sessions, AI answer mentions, visible citations, recommendations, qualified leads, and revenue are different events. They belong in one funnel but should not be collapsed into a single score without showing the underlying counts. Search Console itself documents aggregation and privacy limitations; analytics attribution can also depend on tagging and collection behavior.
4. Use a publication checklist
- Link the original source, not a secondary roundup.
- Record publication and collection dates.
- State sample size, markets, engines, devices, and prompt selection.
- Distinguish repeated observations from unique prompts.
- Report failures and missing data separately from zero.
- Avoid causal language for observational comparisons.
- Label forecasts and vendor estimates.
- Recheck volatile platform facts before each update.
5. Aivius reporting standard
Aivius benchmark pages publish the question set, engine scope, collection date, answer count, observed evidence, and limitations beside the results. Brand scans are snapshots, not universal scores. Competitor values are shown only when they were measured under a comparable protocol; unmeasured values remain blank.
For content performance, review Search Console impressions, clicks, CTR, query and page dimensions; then connect visits to the audit funnel with first-party events. The resulting report should distinguish search demand, observed AI visibility, attributable conversion, and commercial outcome.
Reusable evidence note
Observed in [N] valid responses from [engines], collected [date/time zone], using [prompt selection rule] in [locale]. [F] platform failures were excluded and reported separately. The result describes this sample and does not estimate total market share or guarantee future visibility.
Sources and verification
These sources support the operating constraints in this article. Product behavior and search systems can change, so verify current documentation before implementation.