What Makes AI Assistants Actually Open a Page: Data from 3,997 Search Results
3,997 pages were offered. 356 were opened.
3,997
URLs shown in search results
Claude Code, Jun–Aug 2026
356
Actually opened by the model
Strict attribution, same session
8.9%
Open rate
URL-level, not event-level
When an AI assistant searches the web on your behalf, it gets back a list of results. It then decides which ones to actually open and read. Most of them, it ignores.
We wanted to know what separates the pages that get opened from the ones that get passed over. Not what gets indexed, not what gets summarised, but what makes an AI assistant look at your page when it has ten options and will probably pick two.
So we exported every web citation from a year of AI assistant use across ChatGPT, Claude Code and Gemini, and ran the numbers.
What this sample is, honestly
What actually predicted whether a page got opened
Three things dominated. Everything else was noise.
Your brand being in the query is the biggest lever
| URLs shown | Opened | Open rate | |
|---|---|---|---|
| Query names the publisher | 890 | 168 | 18.9% |
| Query does not | 3,107 | 188 | 6.1% |
When someone asks an AI assistant about your company by name, the pages on your domain are three times more likely to be opened than when they ask a generic question and your page happens to appear. The odds ratio is 3.32, with a session-clustered 95% confidence interval of 2.64 to 5.82.
This is the strongest finding in the data. Most of the conversation about AI visibility is about what to put on your pages. This says the bigger factor is whether people are asking about you in the first place.
The page title decides almost everything else
| Title-query token overlap | URLs | Opened | Open rate |
|---|---|---|---|
| Lowest quartile | 1,159 | 53 | 4.6% |
| Q2 | 868 | 57 | 6.6% |
| Q3 | 1,083 | 92 | 8.5% |
| Highest quartile | 887 | 154 | 17.4% |
Pages whose titles closely matched the words in the query were opened at nearly four times the rate of pages with low overlap. Across the full range, the odds ratio is 12.5 (bootstrap CI 5.4 to 27.2). That is by far the largest effect in the model.
This makes sense once you think about what the model actually sees before it decides to click. It sees a title, a URL, and maybe a snippet. It does not see your word count, your schema markup, or your heading structure. The title is the pitch, and titles that echo the question win.
This post's own title was chosen using this finding. If you searched for something like "what makes AI assistants open a page", the title should be right there in the results.
Position still matters, and the decay is clean
Open rate by search result position
URL-level open rates with 95% Wilson confidence intervals. Ranks 1 to 9 are perfectly monotonic (Spearman rho = −1.00). Rank 10 breaks the pattern.
Rank 10 is elevated because three heavily read regulator domains cluster there. Excluding them, the bump is not significant.
The result in position 1 is opened 20% of the time. By position 9, that drops to 2.8%. The decay is perfectly monotonic across ranks 1 through 9, with an odds ratio of 0.78 per position (bootstrap CI 0.749 to 0.836). Rank 1 is opened seven times as often as rank 9.
Rank 10 breaks the pattern at 7.3%. This is not a mysterious preference for the last result. Three heavily cited regulator domains (the FCA and similar) cluster at rank 10 in this dataset. Exclude them and the elevation disappears. We are showing it rather than hiding it because omitting it would look like cherry-picking to anyone who downloads the data.
So traditional SEO ranking still matters for AI visibility. A page that ranks first in the search results an AI runs is seven times more likely to be read than one that ranks ninth.
What did not predict whether a page got opened
We crawled 1,791 of the pages in the dataset and measured the things GEO guides tell you to optimise. None of them showed a detectable effect on whether the AI opened the page.
| Factor | Opened pages | Skipped pages | p-value |
|---|---|---|---|
| Word count, median | 1,988 | 2,039 | 0.46 |
| Number of headings, median | 16 | 17 | 0.54 |
| Question-shaped headings, share | 10% | 11% | 0.91 |
| FAQ schema present | 19.2% | 20.0% | 0.80 |
| Any schema present | 62.0% | 68.9% | 0.03 |
| Visible publication date | 55.7% | 61.5% | 0.08 |
| Median page age | 251 days | 194 days | 0.91 |
You might have noticed this post does not carry FAQ schema. That is deliberate. Our own data showed no detectable effect from it (p = 0.80), so it would be odd to stuff the post with it. We are using Article schema for provenance, which is honest and sensible.
Blog posts are the second-worst performing format
Open rate by page type
URL-level rates across all 3,997 search results. Controlling for rank and entity match, regulator pages (OR 2.84) and academic pages (OR 3.65) are significantly above the blog baseline. Vendor docs and help centres lose their advantage once entity match is controlled.
Government and regulator pages are opened at more than four times the rate of blog posts, even after controlling for other factors. Academic papers are opened at close to three times the rate.
Vendor documentation and help centres look good in the raw numbers (11.5% and 15.2%), but their advantage disappears once you control for entity match. Their pages were opened because the query named the vendor, not because the format is special.
The blog post, the format most businesses produce for AI visibility, is the second-worst performing page type in the data. Only directory and review sites did worse.
Everyone is blocking the wrong bot
We audited robots.txt across 3,139 domains. Of the 2,736 that serve one, here is what they are blocking.
| Bot | What it does | Domains blocking |
|---|---|---|
| CCBot | Collects training data | 250 (9.1%) |
| GPTBot | Collects training data for OpenAI | 222 (8.1%) |
| ClaudeBot | Collects training data for Anthropic | 207 (7.6%) |
| Google-Extended | Collects training data for Google | 203 (7.4%) |
| PerplexityBot | Indexes for Perplexity search | 74 (2.7%) |
| ChatGPT-User | Reads pages in a live conversation | 49 (1.8%) |
| Claude-User | Reads pages in a live conversation | 36 (1.3%) |
| OAI-SearchBot | Indexes for ChatGPT search | 25 (0.9%) |
Six times as many sites block the training crawler (GPTBot) as block the one that actually reads pages during a conversation (ChatGPT-User). These are different bots with different jobs. Blocking GPTBot does not stop ChatGPT from reading your page when someone asks it a question. It stops your page from appearing in the next model's training data.
These are different decisions, and most sites appear to have made only one of them.
Bot roles come from the official documentation: OpenAI's bot directory and Anthropic's crawler page.
Two ways to be invisible that you would never notice
When our crawler visited domains that appeared in search results but were never opened, it found two distinct failure modes.
Hard-blocked by a WAF. Sites like medium.com, capterra.com and researchgate.net returned HTTP 403. Here is the twist: capterra.com's robots.txt explicitly allows every major AI bot. The 403 came from a web application firewall reacting to an unfamiliar user agent, not from an AI access policy. Your bot-protection layer can override your stated AI policy without anyone noticing.
Returns 200 but empty. link.springer.com returned HTTP 200 with a median of 37 words. That is a bot wall dressed as a successful response. Any monitoring that only checks status codes scores this as perfectly healthy.
Both are things you can check from the outside, and both would be invisible to standard uptime monitoring.
The long tail is real
The top 15 domains account for 25.7% of ChatGPT citation events. Eighty-two domains account for half. And 1,034 domains contributed only a single URL.
The distribution is long-tailed, not winner-take-all. Small, focused sites do appear in AI search results. If your page is relevant to the query and the title matches, the assistant will find it regardless of your domain authority.
Get the AI citation playbook
The findings above turned into actionable guidance: which bots to unblock, how to audit what crawlers actually receive, title and entity recommendations, and what you can measure in GA4 today. One PDF, no fluff.
How we did this
The dataset. We exported citations from ChatGPT (11,208 verified events, Oct 2023 to Aug 2026), Claude Code (5,872 rows, Jun to Aug 2026) and checked 3,139 unique domains. The Claude Code data is the most valuable half because it records the full search result set with rank positions, giving us negative evidence: pages that were offered and ignored. We have not found another public citation study that includes that.
Read definition. A URL counts as opened only if the assistant fetched it in the same session, after the search that showed it, within 60 minutes. An earlier, looser definition (fetched anywhere in the corpus) inflated every rate. 44% of read credits under the loose definition were cross-session or backdated. We use the strict definition throughout.
Statistics. Logistic regression on the full population of 3,997 URLs, with session-clustered bootstrap confidence intervals (300 repetitions) and Benjamini-Hochberg correction for multiple comparisons across domain tests.
Adversarial verification. Before writing anything, two independent audits challenged every finding. They caught a bug in read attribution and a unit-mixing error in domain-level tests. Four of the six original findings were withdrawn as a result. What survived is what you have read here. The analysis broke most of its own findings before we published, and we are telling you that upfront.
The crawl. 2,406 URLs were profiled for page structure, respecting robots.txt and rate-limited to one request per host every 1.5 seconds. Where robots.txt disallowed a generic crawler, the disallow was recorded as data and never bypassed. 74.4% of URLs returned content. Crawl failure was not differential by outcome (p = 0.56).
Download the full dataset (CSV, 3,997 rows)
Cite this study
If you reference this data, please credit:
Rodell, A. (2026). "What Makes AI Assistants Actually Open a Page: Data from 3,997 Search Results." Qwestyon. https://www.qwestyon.com/blog/what-makes-ai-assistants-open-a-page
The dataset and methodology are published so you can verify, extend, or challenge any finding. The more people look at this data the better.
Written by
Adam Rodell
Founder of Qwestyon, a Brighton-based digital marketing studio. Adam runs paid search and paid social for UK businesses and publishes original research on Generative Engine Optimisation, including the schema and AI-search audits on this blog.