How to track AI search
engine citations.
Fix the platforms, prompts and window first. Record every source reference the same way, keep citations apart from mentions, keep platform totals. Then read the trend next to rankings, clicks and sessions, because a citation on its own is not traffic.
A trailing number,
not a running total.
The figure is trailing: it shows what the prompt set finds now, so it falls as well as rises. One number hides that. The series is the honest artefact.
Swipe to view the full chart.
Ahrefs Brand Radar · US Brand Overview · all platforms · month-end · exported 11 Sep 2026. Shaded band marks the drawdown.
Two things matter more than the endpoint. The −23% drawdown (27 December to 19 February) came with no change to the pages: the platforms moved. And the peak sits ahead of the current value, which is why a dated figure beats a headline.
Where the citations sit (prompt-only subset, 11 September 2026, 570): AI Overviews 167, AI Mode 152, Gemini 82, Copilot 59, ChatGPT 55, Perplexity 55. Google Search’s AI Overviews and AI Mode carry 56%.
Rankings improved while
click-through fell.
On the US blog, average position improved from 14.9 to 9.5 while click-through fell 43%. A group of top-ten pages drew heavy impressions at under 0.5% CTR. Visibility was being absorbed before the click.
The response was an AEO/GEO programme: clear entities, extractable answers, evidence, structure, and this tracking. The organic-search evidence is in the roadsurfer case study.
How to monitor AI search
visibility.
Run a stable set of prompts across the platforms your audience uses. For each, note whether the brand is cited, only mentioned, or absent, and which URL. Read the trend beside search performance.
Stability beats volume. Change the prompts, platforms or definitions mid-window and a rising total means less than it looks.
How to track brand
mentions in AI search.
Use prompts that describe the category and the buying situation without naming the brand. Log linked citations and unlinked mentions separately, note which competitors appear, and let the gaps set the content priorities.
What I would, and would not,
claim.
Questions to ask of
any citation number.
01How to track AI search engine citations
Fix the platforms, topics, prompt set and reporting window before you collect anything. Record every visible source reference the same way, keep citations separate from mentions, and keep platform-level totals so a coverage change is visible later. Then read the trend beside rankings, impressions, clicks and sessions.
02How to monitor AI search visibility
Run a stable set of commercial and informational prompts across the platforms your audience uses, and record whether the brand is cited, only mentioned, or absent, plus the cited URL. Change the prompt set, platform list or definitions as rarely as possible; when you must, date the change in the series.
03How to track brand mentions in AI search
Use prompts that describe the category, problem and buying situation without naming the brand. Log linked citations and unlinked mentions in separate columns, note which competitors appear in the same answers, and use the gaps to decide where clearer evidence or a better source page is needed.
04What is an AI search citation?
A visible source reference to one of your URLs inside an AI-generated answer, on ChatGPT, Google AI Overviews, AI Mode, Gemini, Perplexity or Copilot. It differs from a mention, where the brand is named in the answer text without a link. In this report, 1,358 counts citations; mentions on 11 September 2026 were 237.
05Does a citation count mean visits?
No. A citation is a visibility event inside an answer, and most answers end without a click. On one new roadsurfer page, nine AI-assistant users arrived in the first three weeks against 1,909 organic users. Citations belong beside traffic metrics, not in place of them.
06Did the programme grow from zero citations?
Not claimed. The report reads 28 at the end of April 2025, where the export starts; the prompt-only subset records its first citation on 24 June 2025. Tracking started after the work did, so the earliest state is a first reading, not a baseline of zero.