Why this dataset exists
Most advice on getting cited by AI is opinion. We had an unusually good laboratory: hardly anyone reaches Appscribed from a search engine, yet AI assistants quote us thousands of times a day. So we queried the assistants directly, at scale, from two locations, and archived every response with a timestamp. The aggregate figures below are published openly so others can reference and reproduce the direction of the findings.
The gap: citations are the new traffic
Across a 91-day window, AI assistants cited Appscribed about 2,160 times a day while sending roughly 50 times fewer web clicks. If you only watch your analytics dashboard, you miss where your buyers actually see you.
Most AI-citation tools undercount you
We first tested a buyer question through the popular API method: an automated tool querying the model 30 times. It cited us 0 out of 30. Then we asked the same question in the real ChatGPT app, the way a person does. It cited us half the time. API tools run from a fixed location on a stripped-down search stack, so they measure a ghost.
Citation is probabilistic, not a single check
We fired one identical question into a clean, logged-out ChatGPT ten times. It cited Appscribed in five of them, scattered rather than clustered, so it is not caching. ChatGPT rewrites your question into different searches each time. A single check tells you almost nothing; a rate tells you everything.
Key findings
Citations, not clicks
AI assistants cited Appscribed 196,554 times in 91 days (about 2,160 a day) while sending roughly 50x fewer web clicks.Rank is not citation
Pages ranking #80 or lower in classic search were still cited. AI retrieval draws on its own index and query fan-out, not the top ten.Comparisons win
Assistants cite third-party comparison and round-up pages for evaluative claims, and vendor pages for product facts. The "best X" page gets quoted, not your home page.Citation is local
The same question surfaced a different roster of sources depending on the searcher's location, so a single blended score hides whether your buyer sees you.A quotable sentence is the surface
AI reaches for a clear, sourced comparative judgment it can safely lift. Pages that state one give the assistant something to quote.Measure a rate
Because citation is probabilistic, fire each buyer question 10+ times, per engine and per market, in the real app, and track the rate over time.
The dataset
The core measured figures, also available as findings.csv:
| Metric | Value | Window / method |
|---|---|---|
| Total AI citations (Google + Bing) | 196,554 | 91-day census, GSC + Bing Webmaster Tools |
| Average citations per day | ~2,160 | derived from the 91-day window |
| Web clicks vs citations | ~50x fewer | GA4, bot-filtered |
| Citation rate, live ChatGPT app | 50% | same query, repeated fires |
| Citation rate, automated API | 0 / 30 | same query, 30 runs, native web |
| Logged-out ChatGPT repeat fires cited | 5 / 10 | clean session, not cached |
| Lowest classic rank still cited | #80+ | observed in retrieval |
Figures are from a fixed measurement window and will drift over time. The direction of each finding is robust; treat the precise percentages as indicative and re-measure your own category.
How we measured it
No single tool sees the whole picture, so we triangulated across five instruments and archived every response the moment we pulled it.
Consumer-app probes
ChatGPT, logged out. Ground-truth citation rate per query and market, fired 10+ times each, archived.
OpenRouter API
The automated-tool baseline on native web. 30 runs per query, saved as raw JSON.
DataForSEO
Live SERP rank, category demand, and a reverse-lookup of competitor citations.
Bing + Google dashboards
The 91-day AI-citation census across both engines: 196,554 citations from Bing Webmaster Tools and Google Search Console.
GA4
AI-assistant referral clicks and engagement, bot-filtered.
The seven moves that earn a citation
Write the comparison, not the brochure
Answer "what's the best X?" better than anyone, on one page.Put a quotable verdict in a sentence
State a clear comparative judgment; that is the sentence an assistant lifts.Lead with evidence
Original data and statistics are quoted disproportionately.Structure it for extraction
Question headings, a 30-to-50-word answer capsule beneath each, tables and lists.Stop treating rank as the goal
You can be cited from page eight, so optimize the page to be quoted, not only ranked.Measure a rate, per engine, per market
One check is noise; fire each query 10+ times and track the rate.Audit the app, not the API
API tools measure the wrong stack and undercount you.
The full study walks through each move with the archived responses and screenshots.
FAQ
Is this the same as SEO?
Overlapping but not identical. SEO earns a ranked link a person clicks; answer engine optimization earns a citation inside the AI's answer. In this data we are cited on pages that rank #80 or nowhere, and we rank on pages AI ignores, so treat it as a second surface to optimize.
How do I check whether my brand is cited by AI?
Fire your buyer's real questions into the live app 10 or more times, from your buyer's location, on each engine, and count the rate. Do not trust a single check or an API tool: here the same query was cited 50% of the time in the app and 0 of 30 via API.
Do I need a big brand or lots of backlinks?
Less than you would for SEO. Assistants weight evidence and consensus over domain size: original data, a quotable verdict, and third-party mentions. A mid-size directory is cited ~2,200 times a day here because the content is the best evidence, not because the brand is the biggest.
Can I reuse this dataset?
Yes. The figures on this page and in findings.csv are published for reference. Please attribute Appscribed and link to the full study.