Findings · 2026-10-30 · By VSNARY | Emmanuel Orta

AI citations on two dates: seven matched questions

The same seven questions on Google AI Mode and Perplexity, asked on 5 and 29 October 2026: four pairs became cited, none were lost. A wider 27-question pass is alongside. Neither shows why.

We asked the same seven questions on Google AI Mode and Perplexity on 5 October and again on 29 October 2026, and recorded whether each answer cited CrawlCheck. Of the fourteen question-surface pairs, four went from not cited to cited and none went the other way. A wider pass on 29 October, 27 questions on both surfaces, recorded CrawlCheck cited in 4 of 27 AI Mode answers and 9 of 27 Perplexity answers. Both are archived manual notes from signed-in sessions on questions we picked ourselves, and neither shows why anything changed.

What changed on the seven matched questions? #

The match is the point. The 29 October pass had 27 questions and the 5 October pass had seven, so comparing totals would mix new questions with changes on old ones. Matching the exact seven question strings first means a new citation is a new citation of a question we had already asked. Publishing the pairs, not only the totals, also shows that the four gains are four different questions, not one question counted twice.

What did the wider 27-question pass record? #

The panel is the seven recurring questions plus twenty newer ones taken from guide titles, mostly about crawling, technical checks or the product itself. It is not a panel of local-service recommendations and should not be read as one. A link in the answer or source list counted as cited; a mention without a link did not. Source lists were summarized by hand, sometimes several links under one label, so they cannot support a count of unique publishers or backlinks.

What these numbers cannot show #

• Cause. The site changed, the products changed, and personalization in signed-in sessions may have changed. There was no holdout and no change log tied to the pairs, so the four flips cannot be credited to any edit. Access repairs have the same limit: see what an access repair proves.
• A trend. Two dates twenty-four days apart cannot tell a trend from a step, or a step from noise on a base of seven. A third date could reverse any of the four.
• That nothing was lost. Only one pair was cited on 5 October, so there was almost nothing to lose. “None flipped off” is true and weak.
• Which engine is better. Nine against four describes this panel on one afternoon. Moving two questions across the line would change the story without changing the design.
• Your share of a category. The questions were chosen around our own content. A logged-out user in another city may have seen something else.

How to run the same check on your own site #

1. Write the questions down before you ask any of them, and keep the wording byte-identical between dates.
2. Record the surface, whether the session was signed in, and the location context.
3. Save the answer and its source list, not only your verdict.
4. Mark cited or not cited against a rule you can reapply, such as a source chip naming your domain, and keep a mention without a link separate from a linked citation.
5. On the next date, match the same questions before counting. New questions are a new panel, never part of a “change” column.
6. Publish the pairs and the question list. A citation rate without the questions is a number nobody can audit. Treat a failed collection as missing data, not lost visibility.

For the wider method, including why an API reading is not what a customer sees, see how to measure AI visibility yourself and AI visibility scores measure an API, not ChatGPT.

Evidence and reuse #

The dated evidence bundle holds the question list, both labels for every pair, the calculations and source-capture notes. These are archived observation labels, not answers replayed for this note, and not a controlled trial of AI ranking. Tables may be reused with attribution to CrawlCheck and this note dated 30 October 2026; keep the sample and limits when quoting a result. Google’s own description of these surfaces is in its AI features documentation.

Questions this post answers

What changed in CrawlCheck's AI citations between 5 and 29 October 2026?

On seven questions asked on both dates, four of fourteen question-surface pairs moved from not cited to cited and none moved the other way. Google AI Mode went from none of seven to two of seven, and Perplexity from one of seven to three of seven.

How often was CrawlCheck cited on the 27-question pass?

On 29 October 2026 CrawlCheck was labeled cited in 4 of 27 Google AI Mode answers and 9 of 27 Perplexity answers. The questions were chosen around CrawlCheck's own content and asked in signed-in sessions, so the counts describe that panel on that day.

Do two snapshots show that a site change caused more citations?

No. Two dates cannot separate a trend from a step or from ordinary answer variation, and there was no holdout group or change log tied to the questions. The site, the AI products and personalization could all have changed between the dates.

Why compare only the seven matched questions?

Because the later pass asked more questions. Comparing totals would count new questions as gains. Matching the exact question text first means any change is a change on a question that was already being asked.

How can I track whether AI answers cite my site?

Fix the question list and wording in advance, record the surface and signed-in state, save each answer and source list, mark cited against a rule you can reapply, match the same questions on every later date and publish the pairs alongside the totals.

Related findings

Comments

Comments are read before they appear. Nothing is published automatically, and no account is needed.

Writing about this? Facts, live figures and marks — every number on that page is dated and traceable to a scan.

All findings · The dataset · How the dataset works