Baseline and retest

A baseline retest means measuring once with a fixed question set before you optimize anything, then repeating the measurement later using exactly the same method, and judging change by the difference between the two.

AI answers are not stable outputs. The same question can produce different answers at different times, from different accounts, and with web search on or off — and the models themselves are updated continuously. Without a baseline, any single measurement is an isolated number that says nothing about whether your work had an effect.

If you come from SEO, think of it as rank tracking with a stricter protocol. A rank tracker checks the same keywords in the same location every day. A baseline retest does the same for AI answers, but has to control more variables, because a generated answer varies far more than a results page.

The three things you must hold fixed

• The question set. Question wording and numbering do not change. New questions are counted separately and are never merged into historical totals.
• The platform list. Record each AI search platform separately. Don’t compute a weighted average across platforms.
• The sampling method. How many times each question is sampled, whether web search is on, which account and region are used, and the collection date.

Change any of these and you have a new baseline, not a retest.

What to record each round

• Raw answer text and any cited URLs.
• The metrics derived from them, such as mention rate, answer position and first recommendation rate.
• The number of valid samples, plus missing samples and why they failed. A question that returned an error is “not measured,” not “not mentioned.”
• What happened between rounds: content going live, third-party coverage, announced platform model updates. This log is what you use later to reason about causes.

How often to retest

Content and site changes only take effect after they have been re-crawled and re-indexed. If the interval is too short, most of what you measure is random variation. A common approach is to take the baseline, finish the changes, and then retest on a fixed schedule — not ad hoc whenever someone wants a number.

Before each retest, confirm that the changes you are testing for are actually live. A retest run against a site that hasn’t deployed its fixes measures nothing useful.

Limits

A retest is an observation, not a controlled experiment. A change in the metrics may come from your own optimization, from a model version update, from a competitor’s activity or from shifts in the underlying content on the web, and these cannot be fully separated.

Conclusions should therefore be written as “under this method, during this period, we observed this change,” not as claims of cause and effect. That phrasing is less exciting, but it’s honest, and it is the only kind of claim the data supports. Be wary of anyone who reports a GEO “result” without saying what the baseline was and how it was measured.

Where it fits

The baseline is usually the first output of an AI visibility audit, and the retest is how you check whether the fixes from that audit made a difference. For more on measurement, see how to measure GEO and how long until GEO works. The free GEO audit starts with exactly this kind of baseline.