Chiang Mai SEO Conference · 13 November 2026
For a year and a half we recorded what five AI engines actually answer, every day, from the real interface. Here is how to measure AI visibility without fooling yourself, which popular proxies don’t hold up, and what actually gets you cited.
What the data says
From Surfer AI Tracker: about 50 million answers from ChatGPT, Perplexity, Gemini, Google AI Overviews and AI Mode.
The playbook
The three acts of the talk as a checklist. Start with measurement: without it you can’t tell a lever from noise.
AI answers are noisy. Some of it you can’t remove; the measurement error you can stop adding yourself.
Two runs of the same prompt share only 40% of brands. After 20 runs, just 4% of brands appear every single time. One run is an anecdote.
Day to day is mostly noise, but answers drift: brand overlap falls from 40% the next day to 28% after 90 days.
Rewording a question keeps about three quarters of its brands. A different intent keeps barely a third. A hundred variants of one question is money spent on the same answer.
Only 29–42% of an API answer’s top-3 brands make the top 3 in the chat people actually use. The API is a different product.
Google’s three surfaces move together; ChatGPT and Perplexity follow nobody. A single “AI visibility” number averages things that don’t belong together.
Because measuring is hard, everyone grabs a proxy. These are the popular ones, checked against citations.
They top global rankings, but only 8.9% of a niche’s core sources come from the global top 20. Find the sources that win in your niche first.
PageRank against citations is flat zero on every engine. Harmonic centrality is the one authority metric that points up, and covering the facts matters about eight times more.
Big sites get their brand named more often, but their pages aren’t cited more. Big isn’t the same as cited.
Old domains win on volume because they have more pages. Inside one prompt, domain age barely matters and even flips sign on ChatGPT and Perplexity.
Every checklist item sits at the bottom of 20 page features. Answering the query early is worth four to five times more.
External links correlate with being cited at 0.03 across all pages and 0.007 in blog posts. That’s noise.
llms.txt flips sign from engine to engine, up on Gemini and down on AI Mode. That’s not a signal.
Schema is up on every engine, mostly because it marks sites someone actually maintains. Keep structured data on pages you look after; skip llms.txt.
What holds up in the data. For the full listicle playbook, see the follow-up talk Who writes the list?
95% of cited domains rank in Google’s top 10 for at least one keyword. SEO isn’t dead; it’s the entry ticket.
The first time your own site gets cited is worth about +25% visibility. Later citations add little on top.
Almost a third of all citations come from the first 50 words of a page. Answer immediately, not after the intro.
A page that answers exactly what was asked is a source 42% of the time; an off-intent blog post, 13%. The strongest single writing signal.
AI quotes passages, not pages. Every paragraph has to make sense on its own.
Coverage drops with every step down the ranking: 39% of the facts in the top 3, 26% by rank 50.
Unique pages win, and what makes them unique isn’t original research. It’s specific, concrete detail.
On someone else’s list your brand is named about half the time at #1 and barely at all at #10. On your own list, put yourself first: 38% instead of 26%.
Brands with four or more cited lists of their own are mentioned in 38% of answers, against 26% with none.
57–74% of citations in AI Mode, AI Overviews, ChatGPT and Perplexity go to pages older than six months. The page fades fast, but the name stays: 83% of the mention survives eleven weeks.
AI instructions
Two audit prompts: one for a page, one for how you measure. Every check comes with the number behind it, and the numbers travel with the prompt, so the model knows what matters most.
Paste a page and the query you want it cited for. The model checks it against eight data-backed checks and tells you what to fix first, and what to stop doing.
Intent: does the page type match what the query asks? A “which” query needs a comparison, a “how” query needs steps.
42% vs 13%how often an on-intent page vs an off-intent blog post is cited
First 50 words: is there a direct, complete answer before any intro?
≈⅓of all citations come from a page’s first 50 words
Passages: does each paragraph name its subject and make sense if quoted alone?
PassagesAI quotes passages, not whole pages
Fact coverage: list what a complete answer needs and mark which facts are missing.
39% → 26%facts covered by top-3 vs rank-50 sources
Concrete detail: are there specific numbers, settings, examples or first-hand results, or only generic claims?
0.12correlation of a unique, concrete angle with citations
Google: does the page rank in the top 10 for the query or close variants? If not, flag it as the first fix.
95%of cited domains rank in Google’s top 10
Structured data: is JSON-LD present and current (Article, Person as author, dateModified)?
+3.7 ppcitation rate with vs without JSON-LD; Product markup +0.1 pp
Wasted effort: flag llms.txt, “trust” links to Wikipedia, and stats, quotes, FAQs or a TOC added only for a checklist. Mark them low priority, not as fixes.
≈ 0llms.txt −0.04 pp without big hosts; external links 0.03; checklist items 4–5× weaker than an early answer
Audit the page below for how likely AI answer engines (ChatGPT, Perplexity, Google AI Overviews and AI Mode, Gemini) are to cite it as a source. The checks come from Surfer AI Tracker data: about 50 million AI answers scraped daily from the real interfaces of five engines. They describe what cited pages have in common; treat them as strong signals, not guarantees. INPUTS - Target query: [the exact question people ask] - Page: [paste the text or the URL content] - Our brand: [Brand] - Google position for the query, if known: [e.g. 7, or "not ranking"] CHECKS (each with the data behind it) 1. Intent: does the page type match what the query asks? A “which” query needs a comparison, a “how” query needs steps. [data: 42% vs 13% how often an on-intent page vs an off-intent blog post is cited] 2. First 50 words: is there a direct, complete answer before any intro? [data: ≈⅓ of all citations come from a page’s first 50 words] 3. Passages: does each paragraph name its subject and make sense if quoted alone? [data: Passages AI quotes passages, not whole pages] 4. Fact coverage: list what a complete answer needs and mark which facts are missing. [data: 39% → 26% facts covered by top-3 vs rank-50 sources] 5. Concrete detail: are there specific numbers, settings, examples or first-hand results, or only generic claims? [data: 0.12 correlation of a unique, concrete angle with citations] 6. Google: does the page rank in the top 10 for the query or close variants? If not, flag it as the first fix. [data: 95% of cited domains rank in Google’s top 10] 7. Structured data: is JSON-LD present and current (Article, Person as author, dateModified)? [data: +3.7 pp citation rate with vs without JSON-LD; Product markup +0.1 pp] 8. Wasted effort: flag llms.txt, “trust” links to Wikipedia, and stats, quotes, FAQs or a TOC added only for a checklist. Mark them low priority, not as fixes. [data: ≈ 0 llms.txt −0.04 pp without big hosts; external links 0.03; checklist items 4–5× weaker than an early answer] OUTPUT 1. A table: check | what you found (quote the page where useful) | pass / partly / fail | the fix. 2. The missing facts from check 4 as a list. 3. A rewritten first 50 words that answer the query directly. 4. The 5 fixes that matter most, in order of impact. 5. What to stop spending time on (from check 8), if anything.
Before you trust any AI visibility number. Describe how you collect answers and the model grades it against eight rules.
Of the brands you see across several daily runs of the same prompt, this share shows up in every one of them.
Run every prompt many times; never judge a brand from a single answer.
40%of brands are shared by two runs of the same prompt
Track every day, but read trends over weeks.
40% → 28%brand overlap one day vs 90 days later
Expect sources to churn faster than brands.
37% → 13%cited-domain overlap one day vs 90 days later
Track different intents, not rewordings of one question.
−8 vs −19 ppbrand overlap lost by rewording vs by asking another engine
Collect answers from the real chat interface, not the API.
29–42%of API top-3 brands make the chat’s top 3 (chat vs itself: 49–54%)
Report each engine separately, never one blended score.
0.32–0.40how closely ChatGPT and Perplexity track other engines (Google surfaces: 0.58–0.70)
Don’t use domain authority as a stand-in for AI visibility.
≈ 0correlation of PageRank and Domain Score with citations, on every engine
Benchmark against your niche’s sources, not global top domains.
8.9%of a niche’s core sources come from the global top 20
Review my AI visibility tracking setup against measurement rules from Surfer AI Tracker data (about 50 million AI answers from five engines, scraped from the real interfaces). INPUTS - Tool or method: [how answers are collected] - Engines: [list] - Prompts: [how many, how they were chosen] - Runs per prompt and frequency: [e.g. once a week] - Reporting: [one score, per engine, per prompt?] RULES (each with the data behind it) 1. Run every prompt many times; never judge a brand from a single answer. [data: 40% of brands are shared by two runs of the same prompt] 2. Track every day, but read trends over weeks. [data: 40% → 28% brand overlap one day vs 90 days later] 3. Expect sources to churn faster than brands. [data: 37% → 13% cited-domain overlap one day vs 90 days later] 4. Track different intents, not rewordings of one question. [data: −8 vs −19 pp brand overlap lost by rewording vs by asking another engine] 5. Collect answers from the real chat interface, not the API. [data: 29–42% of API top-3 brands make the chat’s top 3 (chat vs itself: 49–54%)] 6. Report each engine separately, never one blended score. [data: 0.32–0.40 how closely ChatGPT and Perplexity track other engines (Google surfaces: 0.58–0.70)] 7. Don’t use domain authority as a stand-in for AI visibility. [data: ≈ 0 correlation of PageRank and Domain Score with citations, on every engine] 8. Benchmark against your niche’s sources, not global top domains. [data: 8.9% of a niche’s core sources come from the global top 20] For each rule: say pass / partly / fail, why, and the cheapest fix. Then list the conclusions I’ve drawn so far that I shouldn’t trust yet, and what to change first.
Method
All data comes from Surfer AI Tracker: about 100,000 prompts a day, scraped since March 2025 from the real interfaces of ChatGPT, Perplexity, Gemini, Google AI Overviews and AI Mode, not from APIs. Each slide lists its own sample and period in the footer.
Correlations and comparisons within prompts and niches, not proven causes. Test on your own pages and prompts before you roll anything out.
2027 is open, with the first events confirmed in March.