For the four weeks from October 20 to November 16, 2026, we tallied each opening of the 25,672 startup profiles on StartupHub.ai by an AI assistant. Claude accounted for most of that attention with 37,748 live reads compared with 4,976 from ChatGPT. The assistants did not focus on famous companies. A startup with a round announced in the last 90 days was read more than four times as often as a startup with no round on record.
How many startup profiles AI assistants read
A live read covers the case that counts for a founder. A person puts a question to an assistant and the assistant pulls a page to form the answer. On that basis, 11,536 of the 25,672 profiles, or 44.9%, were read at least once in the four weeks, with 42,725 reads in total. When every AI agent is counted, including the crawlers that build search indexes, coverage grows to 22,533 profiles, or 87.8%.
Fetching is dominated by those crawlers. ChatGPT's search crawler and Perplexity's crawler fetched startup profiles 313,924 times together, or about seven fetches for every live read.
Claude reads the page, ChatGPT reads its copy
Claude opened 7.6 times as many profile pages as ChatGPT. The difference does not show which assistant serves more users. The difference shows how each assistant answers. The ChatGPT search crawler fetched profiles 150,392 times in the same four weeks, or 30 times more often than ChatGPT called one up live for a user, so most ChatGPT answers about a startup draw on a copy already stored. Claude reads the page while the person waits, with 37,748 reads across 10,714 different profiles.
Perplexity is the missing piece in this study. Its crawler ran hardest at 163,532 profile fetches, yet its live reader almost never shows under its own name, so we omit Perplexity's live reading instead of estimating it.
Pages, not llms.txt
The two AI assistants also read different material. Over the last seven days, 84% of live reads from ChatGPT on StartupHub.ai were news articles and 5.5% were startup profiles. Reads from Claude spread out, with 32% articles, 23% profiles, 12% calls to the API and 10% markdown copies of pages, the plain-text versions published for agents.
ChatGPT requested a markdown copy 19 times in that week, while Claude made 3,874 such requests. Neither AI assistant requested llms.txt once. The file proposed as a map of a site for AI assistants drew no live requests from either system in seven days.
Which startups get looked up
Funding history marks the sharpest split. Among profiles that list no funding, 32% got a live read. For companies that had raised any amount the share rises to between 60% and 64%, and that range held from seed companies under $10 million up to those that had raised $1 billion.
Above that level the rate drops. Of the 475 companies that had raised more than $1 billion, only half were read live in the four weeks, and those reads were less frequent at 1.9 times on average compared with about 2.6 for the companies just below. Assistants already carry knowledge of very large companies from training and tend to look up the ones they do not.
| Total funding on record | Profiles | Read live | Live reads (avg) |
|---|---|---|---|
| None on record | 14,653 | 32.1% | 0.97 |
| Under $10M | 3,856 | 60% | 2.59 |
| $10M to $100M | 4,520 | 64.2% | 2.67 |
| $100M to $1B | 2,169 | 63.8% | 2.56 |
| Over $1B | 475 | 50.1% | 1.89 |
Recency counts alongside scale. Startups that closed their most recent round within the past 90 days drew 4.2 live reads on average. The figure fell to 2.2 for startups whose previous financing dated back over a year. Where StartupHub.ai holds no round entry, the average was 1.0. A fresh raise puts founders under research.
Company age points the same way. Firms founded two to five years ago drew the most live attention, with 62% read live and an average of 3.2 times. Younger firms founded in the last two years were read less often at 38%, and firms older than ten years were also lower at 41%. The middle period is when buyers and investors are still making up their minds.
| Company age | Profiles | Read live | Live reads (avg) |
|---|---|---|---|
| Under 2 years | 2,961 | 38.2% | 1.97 |
| 2 to 5 years | 3,933 | 62.3% | 3.16 |
| 5 to 10 years | 7,087 | 55.5% | 1.85 |
| Over 10 years | 7,324 | 40.9% | 1.12 |
Thousands of companies, a couple of reads each
Live reading stayed unusually broad. The top 1% of profiles drew 9.3% of all live reads, while the top 10% of profiles drew 39%. A profile that was read at all got two reads in the typical case, and only two profiles passed 100 reads. Assistants are not pounding a small set of famous names. Assistants are looking up thousands of companies a couple of times apiece.
Many of those companies drew no other visitors. Among the 18,836 profiles with no recorded human visit in the four weeks, 41% were still read by an assistant for someone. Across much of the startup world, software working for a person now becomes the first reader of a company page.
Are the readers real?
A user agent works as a name tag that any requester can imitate, so StartupHub.ai traced the origin addresses instead. In the examined request sample, 62 of 63 requests labelled ChatGPT arrived from OpenAI's published address ranges for its live reader. Of 1,124 requests labelled Claude, 634 originated on Anthropic's own network while the remainder used home and office connections. That second pattern matches how Claude Code fetches a page, from the user's own machine. The Claude desktop app fetches a page the same way. That check showed no evidence of scanners borrowing the names.
What happens when an agent tries to use the site
Reading a page is one thing. Doing something with it is another, and it depends on the agent as much as the model. Since October 1, our Live Agent Run has pointed five real agent harnesses, Claude Code, Codex CLI, OpenCode, Cursor and Gemini CLI, at 728 websites, 4,780 runs in all, each asked the same questions about the site, with a plain model and a single fetch tool alongside as the control. Gemini CLI, the slowest of the five, had run on 59 of them when this was published.
The plain model found 2.5 answers on average from 1.3 fetches. The harnesses found between 3.8 and 6.0 answers, from 5.0 to 8.3 fetches. Same questions, same sites: the harness decided how much of each site the agent reached.
What founders can do
- Make sure funding is documented publicly. About 60% of startups showing funding on record got a live read, compared with 32% of those without.
- Announce a recent raise where assistants check. For companies with a round in the last 90 days, the average was 4.2 live reads, four times the rate of those with none on record.
- Offer a markdown version of key pages with a link from the page itself. In a week on StartupHub.ai alone, Claude requested markdown copies 3,874 times.
- Do not rely on llms.txt alone. Neither assistant requested it in seven days.
- See what an agent can actually reach on your site with the free Agent Readiness scan.
How we measured
StartupHub.ai counted a read as a request for a startup profile page from October 20 to November 16, 2026. Identification used the labels the assistant makers disclose for live reading systems (Claude-User, ChatGPT-User) plus search crawling systems (OAI-SearchBot, PerplexityBot). Only requests that returned a successful response were included.
The denominator of 25,672 equals all startup profiles published on StartupHub.ai. Counts of human visits rely on StartupHub.ai page analytics. That tool misses people who block scripts, so "no recorded human visit" marks a ceiling on the true count.
The split showing what each assistant requested reflects the seven days ending November 17. A sample of requests was matched to address ranges OpenAI discloses and to Anthropic's network. Numbers for Live Agent Run include all runs from October 1, 2026 through publication day.
Questions
Does ChatGPT read websites in real time?
Partly. During four weeks of user answers, ChatGPT fetched live StartupHub.ai startup profiles on 4,976 occasions, while its search crawler fetched them on 150,392 occasions, about 30 times as often. Most answers from ChatGPT about a company draw on a copy fetched earlier.
Do AI assistants read llms.txt?
That pattern does not show in our data. During seven days, neither Claude nor ChatGPT requested llms.txt from StartupHub.ai even once. Claude pulled our pages in markdown form 3,874 times, which suggests linked plain-text alternatives get far heavier use than the file intended to index them.
Which startups do AI assistants look up most?
Growth-stage businesses with backing draw the attention. Between 60% to 64% of all funded startups got a live read during four weeks, compared with 32% lacking any funding on record. Startups raising in the last 90 days drew 4.2 live reads, four times the rate for those with no round.
How do I get my startup read by AI assistants?
Put assistant-ready facts where they are easy to parse: funding on record, plus a recent announcement, plus a plain markdown copy of the key pages linked from each page. Then see what an agent can actually reach using the free Agent Readiness scan.





