Measuring AI search: a September 2026 retrospective

By Sumith Parambat DamodaranPublished in TechnologyOctober 08, 20264 min read

An AI visibility report is a sample of answers under recorded conditions. It is not a census of what every customer sees. Treating that distinction seriously changes both the dashboard and the decisions made from it.

Retrospective theme date: 9 September 2026. First published: 8 October 2026. This is a newly written reflection organised as part of a July–October series; it was not published on the earlier theme date. Sources and product information were reviewed on 8 October 2026.

Separate the signals

In the BrightonSEO workshop, Myriam Jessier distinguished citations, brand mentions and business outcomes. I would track those separately. A linked citation can provide a route to a page; a brand mention can appear without a link. Either may be positive, inaccurate or irrelevant to a buying decision.

For my own reporting convention, a mention means any answer that names the brand, whether or not it includes a citation. A citation means an answer with a source link to the tracked domain. These categories overlap. Define that explicitly because tools and teams may use different conventions.

Consider a fictional sample of 40 answers. If 12 mention the brand and 5 link to its domain, the sample mention rate is 30% and the sample citation rate is 12.5%. These are illustrative figures, not research or customer results. They describe only those 40 observations.

Record how the sample was produced

Keep the prompt text, platform, collection time, market, language and available configuration with each result. Record whether the test used a consumer interface or an API, whether search was enabled and whether previous conversation context was present. Some providers do not expose an exact model version; report that limitation rather than inventing one.

Choose prompts around decisions that matter. Include discovery, comparison, suitability and factual checks. Group them by intent so a change in one area does not disappear inside an overall average.

Repeat a stable set of prompts to observe variation. If the test conditions change, annotate the report. A jump after switching platforms or adding branded questions cannot fairly be described as an improvement caused by new content.

Inspect accuracy as well as presence

An answer that recommends a product for an unsupported use case is not a win. For important answers, review the claim, the cited page and whether that page actually supports the claim.

My proposed issue log contains the inaccurate statement, supporting screenshot or saved response, affected decision, authoritative correction and owner. This creates a route from measurement to action. A score without a responsible owner tends to become a recurring slide rather than a useful intervention.

Connect cautiously to business outcomes

Referral analytics and customer-reported discovery can add evidence, but attribution has gaps. A person may use an assistant, search for the brand later and arrive directly. Conversely, a recorded AI referral does not establish that the assistant caused a purchase.

Google introduced dedicated generative AI performance reports in Search Console in June 2026 and states they were rolled out worldwide by 31 August. Those reports concern Google’s surfaces; they do not measure all AI assistants. Review them alongside other evidence rather than combining unlike metrics into a universal visibility score.

Run a modest experiment

Select one group of decision pages, record a baseline and publish a specific correction. Keep a comparable group unchanged where practical. Repeat observations and examine both factual accuracy and useful site behaviour. Seasonality and platform changes remain alternative explanations.

The workshop’s warning about controlled prompt tests is the principle I would keep: use them to discover problems and test hypotheses, then seek independent evidence before claiming commercial impact.

Sources and attribution

Myriam Jessier, Fundamentals of GEO, BrightonSEO 2026 workshop, supplied slide deck, slides 36–39. Metric definitions here are explicitly my reporting conventions.
Google: generative AI performance reports in Search Console, checked 8 October 2026.

Illustrative cover image: existing site photograph by Pankaj Patel, reused from the Flutter VSCode extensions article. It is not workshop or product imagery.