What's actually been measured about AI search — not forecast
Every claim about AI search is either something someone measured or something someone predicted, and almost everything circulating online is the second kind. This page publishes only the first: Pew's browser-instrumented panel on what happens to a click, Cloudflare's crawl-to-referral ratios, and the crawler documentation Google, OpenAI, Anthropic and Perplexity publish themselves. The famous forecasts — the 60%-zero-click-by-2027 kind — are in the dropped section below, with the trail that leads nowhere.
| Metric | Value | Period | Source | Cite |
|---|---|---|---|---|
Click-through rate on a traditional search result link, comparing Google searches that showed an AI summary vs. searches that did not, among a tracked panel of US adults When Google shows an AI summary above your listing, expect roughly half the click-through you'd get on a plain results page — that's a reason to invest in being the cited source inside the summary, not just the blue link below it. This is measured from real, browser-instrumented behavior of about 900 US adults (68,879 Google searches, 12,593 of them with an AI summary), not a survey of what people say they do — it is the strongest evidence available on this question. It covers all Google search categories, not local-business queries specifically. Verbatim from the sourceUsers clicked on a traditional search result link in just 8% of visits to pages with an AI summary, compared with 15% of visits to pages without one. Checked 2026-09-09 | 8% vs. 15%% of visits | March 2025 | Pew Research Center, "Do people click on links in Google AI summaries?" | |
Share of visits to pages with an AI summary where the user clicked a source link embedded inside the summary itself (as opposed to a regular result below it) Don't treat 'getting cited in the AI summary' as equivalent to a click — assume almost nobody follows that specific link, and keep your phone number and booking link visible everywhere else you can be found (Maps, GBP, directories). This is measured among the roughly 12,593 tracked visits to pages that actually showed an AI summary, not across all 68,879 searches in the study — and it still confirms that being cited inside a summary rarely turns into a click, even when the summary is right in front of the user. Verbatim from the sourceThis occurred in just 1% of all visits to pages with such a summary. Checked 2026-09-09 | 1%% of visits to pages with a summary | March 2025 | Pew Research Center, "Do people click on links in Google AI summaries?" | |
Share of search sessions that ended (no further clicks) after a results page showing an AI summary, vs. after a page with only traditional results A meaningful share of your potential customers get their answer and stop right there — that raises the value of your business actually being the fact the AI cites (hours, price range, service area), not just ranking well. Session-ending is the closest measured proxy for a real zero-click search in this dataset; it is still a US-panel average across all search topics, not isolated to local or 'near me' queries. Verbatim from the sourceThis happened on 26% of pages with an AI summary, compared with 16% of pages with only traditional search results. Checked 2026-09-09 | 26% vs. 16%% of visits | March 2025 | Pew Research Center, "Do people click on links in Google AI summaries?" | |
Share of tracked Google searches that produced an AI-generated summary, and share of users who encountered at least one AI summaries are common but not yet universal — most searches still land on a traditional results page, so classic local SEO and Maps visibility are not obsolete. This was still a partial rollout as of March 2025 (18% of queries, not most of them); Google's own later figures further down this page show the feature scaling further afterward, so this specific percentage is dated to that one month. Verbatim from the sourceAbout six-in-ten respondents (58%) conducted at least one Google search in March 2025 that produced an AI-generated summary... Overall, an AI summary appeared in around one-in-five Google searches (18%) that respondents conducted in March 2025. Checked 2026-09-09 | 18% of searches; 58% of users% of searches | March 2025 | Pew Research Center, "Do people click on links in Google AI summaries?" | |
Ratio of pages an AI or search crawler fetched to referral visits it sent back to the same sites, tracked over time for four platforms AI chatbot crawlers read your site at a scale wildly out of proportion to any traffic they send back — don't expect a ChatGPT or Claude citation to move your analytics the way ranking in Google Search does. This is Cloudflare's own network-wide traffic, not a per-industry or per-local-business breakdown. Ratios for OpenAI and Anthropic stay three to five orders of magnitude larger than Google's even after the swings shown here — Google still sends far more visits per page it crawls. See the disagreement below: a separate Cloudflare snapshot from a single week in June puts Anthropic's ratio at roughly 71,000:1, a reminder that any one crawl-to-referral number is a moment, not a constant. Verbatim from the sourceOpenAI ratio dropped slightly, from 1,217 bots per human in January to 1,091 in July (-10%). [...] Anthropic [...] from 286,930 bots per human in January, to 38,065 bots per human in July [...] Perplexity moved in the opposite direction, with bot crawling increasing +256.7% relative to human visitors; climbing from 54 bots per human in January to 195 bots per human in July. [...] Google[']s ratio moved from 3.8 in January to 5.4 in July. Checked 2026-09-09 | Google 3.8→5.4; OpenAI 1,217→1,091; Anthropic 286,930→38,065; Perplexity 54→195 (crawl requests per human referral)crawl requests per referral visit | January–July 2025 | Cloudflare Blog, "The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals" | |
Googlebot's and GPTBot's share of total crawler traffic, and their year-over-year growth in request volume Blocking GPTBot in robots.txt no longer means 'nobody sees me in AI answers' — it means opting out of a crawler now several times larger than a year ago; weigh that against any AI-training privacy concern deliberately, not by default. Measured across Cloudflare's customer network using a fixed set of sites to control for customer growth; this describes crawl volume, not resulting traffic, and is not specific to home-service or local-business sites. Verbatim from the sourceGooglebot, which indexes content for Google Search, was clearly the top crawler throughout the period and showed strong growth, up 96% from May 2024 to May 2025. [...] GPTBot (from OpenAI) increased its share from 2.2% to 7.7% (+5.5 pp), with a 305% rise in requests. Checked 2026-09-09 | Googlebot 30%→50% (+96% requests); GPTBot 2.2%→7.7% (+305% requests)% of crawler traffic | May 2024–May 2025 | Cloudflare Blog, "From Googlebot to GPTBot: Who's crawling your site in 2025" | |
Share of top-10,000 domains (by Cloudflare Radar) whose robots.txt file contains an explicit allow or disallow directive targeting an AI bot, and how many block GPTBot specifically Most large sites still haven't bothered to configure AI-bot rules in robots.txt, so a considered choice here (rather than a default block) is a place a small business can actually be more deliberate than the average big publisher. The universe here is Cloudflare Radar's top 10,000 domains globally, which skews toward large, high-traffic publishers and platforms — most small local-service-business sites have never touched robots.txt at all, so this is not a base rate for that population. Verbatim from the sourceout of 3,816 domains (from the top 10,000) where we were able to find a robots.txt file, 546 (about 14%) had "allow" or "disallow" (fully or partially) directives targeting AI bots. [...] The most frequently blocked bot was GPTBot, disallowed by 312 domains (250 fully, 62 partially). Checked 2026-09-09 | 14% of domains with a robots.txt; GPTBot blocked by 312 domains% of domains | as of 2025-06-06 | Cloudflare Blog, "From Googlebot to GPTBot: Who's crawling your site in 2025" | |
Which crawler retrieves content for Google AI Overviews, and what robots.txt controls apply You can't cherry-pick 'show my page in normal Search but not in AI Overviews' — the only controls are the same ones that govern whether Google can crawl and snippet the page at all. This is Google's own current documentation, not a third party's interpretation. Google states elsewhere on the same page that 'there are no additional requirements to appear in AI Overviews or AI Mode' — confirming there is no way to opt a page into or out of AI Overviews specifically, short of blocking or allowing Googlebot itself (or using nosnippet/noindex on the whole page). Verbatim from the sourceAI is built into Search and integral to how Search functions, which is why robots.txt directives for Googlebot is the control for site owners to manage access to how their sites are crawled for Search. Checked 2026-09-09 | Googlebot (no separate AI Overviews user agent) | current, accessed 2026-09-09 | Google Search Central, "AI features and your website" | |
What the Google-Extended robots.txt token controls, and its stated effect (or non-effect) on Search ranking Blocking Google-Extended is a free, zero-risk-to-rankings way to opt your site's content out of Gemini AI training if you want to — it has no bearing on whether customers find you in ordinary Google Search or Maps. Google-Extended has no HTTP user agent of its own — Google fetches pages with its normal crawler and applies the robots.txt token as a usage flag afterward, so it will never show up as a distinct hit in server access logs. The same page states its purpose is letting publishers manage whether crawled content trains or grounds Gemini models. Verbatim from the sourceGoogle-Extended does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search. Checked 2026-09-09 | Controls Gemini model training/grounding only; explicitly does not affect Search ranking or indexing | current, accessed 2026-09-09 | Google for Developers, "Google's common crawlers" (Crawling infrastructure docs) | |
The distinct purposes of OpenAI's GPTBot (model training) versus OAI-SearchBot (ChatGPT search results), and that they are controlled independently You can block GPTBot (opt out of AI training) while still allowing OAI-SearchBot (stay eligible to be cited in ChatGPT search answers) — these are independent settings, not an all-or-nothing choice. A separate agent, ChatGPT-User, fires only when a live ChatGPT user's query triggers a page fetch, and OpenAI states robots.txt rules 'may not apply' to that one since it is a user-initiated action rather than automated crawling. Verbatim from the sourceGPTBot [...] is used to crawl content that may be used in training our generative AI foundation models. [...] OAI-SearchBot [...] is used to surface websites in search results in ChatGPT's search features. [...] Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links. Checked 2026-09-09 | Two separate crawlers, controlled by separate robots.txt rules | current, accessed 2026-09-09 | OpenAI Developer Platform, "Overview of OpenAI Crawlers" | |
Anthropic's three distinct web crawlers (training, user-triggered fetch, and search indexing) and how each is blocked If you want Claude to keep being able to cite your business when a user asks it a live question, but keep your content out of Anthropic's training data, block ClaudeBot specifically and leave Claude-User and Claude-SearchBot alone. Blocking is per-agent — 'User-agent: ClaudeBot / Disallow: /' stops training crawling only, and does not by itself stop Claude-SearchBot or Claude-User from fetching the page. Verbatim from the sourceClaudeBot helps enhance the utility and safety of our generative AI models by collecting web content that could potentially contribute to their training. [...] Claude-User supports Claude AI users. When individuals ask questions to Claude, it may access websites using a Claude-User agent. [...] Claude-SearchBot navigates the web to improve search result quality for users. Checked 2026-09-09 | ClaudeBot (training), Claude-User (live user fetch), Claude-SearchBot (search index) — each blocked independently | current, accessed 2026-09-09 | Claude Help Center, "Does Anthropic crawl data from the web, and how can site owners block the crawler?" | |
Perplexity's two crawlers — PerplexityBot (search index, honors robots.txt) versus Perplexity-User (live user fetch, generally ignores robots.txt) Disallowing PerplexityBot removes you from Perplexity's pre-built search index, but does not guarantee Perplexity will never fetch and quote your page live when someone asks about your business by name. Because Perplexity-User fetches ignore robots.txt, a page can still be read and quoted to an individual Perplexity user asking about it in real time even if the site's robots.txt fully disallows PerplexityBot from indexing it in advance. Verbatim from the sourcePerplexityBot is designed to surface and link websites in search results on Perplexity. It is not used to crawl content for AI foundation models. [...] Perplexity-User supports user actions within Perplexity. When users ask Perplexity a question, it might visit a web page to help provide an accurate answer. [...] Since a user requested the fetch, this fetcher generally ignores robots.txt rules. Checked 2026-09-09 | PerplexityBot honors robots.txt; Perplexity-User generally does not | current, accessed 2026-09-09 | Perplexity Docs, "Perplexity Crawlers" | |
Total referral visits from AI platforms (ChatGPT, Gemini, Perplexity, Claude, etc.) to the top 1,000 websites globally, compared to referral visits from Google Search, in the same month Even amid triple-digit year-over-year growth, AI-platform referral traffic remains a rounding error next to Google Search referral traffic — for a local business, Google Search and Maps visibility still matters far more than chasing an AI-chatbot citation. This describes the top 1,000 global websites — mostly large publishers, retailers and platforms, not local service businesses. It is Similarweb's clickstream-based estimate, reported here via TechCrunch's coverage of the data rather than Similarweb's own report page, which could not be independently re-verified for these exact figures in this pass; the ratio (about 1 AI referral for every 169 Google referrals) is the useful takeaway, not the absolute counts. Verbatim from the sourceAI platforms in June generated over 1.13 billion referrals to the top 1,000 websites globally... Google Search still accounts for the majority of traffic to these sites, accounting for 191 billion referrals during the same period of June 2025... up 357% since June 2024... ChatGPT dominates here, accounting for more than 80% of the AI referrals to the top 1,000 domains. Checked 2026-09-09 | 1.13 billion (AI) vs. 191 billion (Google Search)billion referral visits | June 2025 | Similarweb data, reported by TechCrunch, "AI referrals to top websites were up 357% year-over-year in June, reaching 1.13B" | |
Google's own stated monthly user count and geographic reach for AI Overviews, as announced on an Alphabet earnings call AI Overviews are no longer a niche experiment your customers might stumble on — at this scale, assume any commercial local-service query has a real chance of surfacing one before your listing. This is self-reported by Google on an investor call with no published methodology or definition of 'monthly user' beyond the figure itself — treat it as a directional scale indicator, not an independently audited measurement. Verbatim from the sourceAI Overviews now has over 2 billion monthly users across more than 200 countries and territories and 40 languages. Checked 2026-09-09 | Over 2 billion monthly users, 200+ countries/territories, 40 languagesbillion monthly users | Q2 2025 (announced 2025-07-23) | Google, "Alphabet Q2 earnings call: CEO's remarks" | |
Google's own claim about how much AI Overviews increases total query volume, for the specific query types where the feature appears More total searches happening does not mean more visits to your site — Google's own framing here is about engagement with Search itself, not about traffic sent onward to businesses. This describes more searches being run, not more clicks reaching websites — it is consistent with, not contradictory to, the Pew finding that each individual search is less likely to produce a click; Google discloses no baseline, comparison period, or methodology for the 10% figure. Verbatim from the sourceWe know how popular AI Overviews are because they are now driving over 10% more queries globally for the types of queries that show them. Checked 2026-09-09 | Over 10% more queries globally% increase in queries | as stated 2025-07-23 | Google, "Alphabet Q2 earnings call: CEO's remarks" | |
Share of AI-chatbot users (specifically for news) across 48 surveyed markets who say they always or often click through from a chatbot's answer to the original news source This number belongs to news publishers, not home-service businesses — don't borrow a 42% click-through figure to justify AI-search investment for a local business; use the Pew and Cloudflare rows above instead, which measure general and crawler-level behavior. This measures news-publisher click-through specifically, from people who already use AI chatbots for news, across a global 48-market self-reported survey (~2,000 respondents per market) — it is not about local-service or transactional queries and should not be applied to a plumber's or dentist's citation rate in a chatbot answer. Verbatim from the source42% of AI chatbot users for news say they always or often click through from chatbot answers to original news sources Checked 2026-09-09 | 42%% of AI-chatbot-for-news users | fieldwork early 2026, published 2026-06-16 | Reuters Institute for the Study of Journalism, Digital News Report 2026 — Executive Summary |
Showing 16 of 16 rows
Where the sources disagree
We do not pick the “right” number for you. Where two honest sources count differently, both are here, along with the reason they differ.
How large the AI-crawler-to-referral imbalance is for OpenAI and Anthropic, and whether it is improving or worsening
- Anthropic ~71,000:1 crawl-to-referral Cloudflare Blog, "The crawl before the fall... of referrals", June 19-26, 2025
- Anthropic 286,930:1 in January 2025, improving to 38,065:1 by July 2025 Cloudflare Blog, "The crawl-to-click gap", January-July 2025
Both numbers are Cloudflare's own data, but drawn from different snapshot windows — a single week versus a six-month trend — and possibly different methodology versions as Cloudflare iterated on its Radar AI Insights product through 2025. The underlying ratio for Anthropic swings by more than an order of magnitude depending on which week is sampled. That swing is the finding, not a flaw to average away: treat any single crawl-to-referral number, from any platform, as a snapshot of one window, not a stable constant you can quote a year later.
What did not survive checking
These are claims we went looking for and could not confirm at the source. They are not in the table — and they are listed here so you do not spend time on them either.
- "By 2027/2028, X% of searches will be zero-click" (commonly cited as 'nearly 60%' or similar, traced to various SEO-industry forecasts building on a 2019 Rand Fishkin/Jumpshot analysis)
- Every version traced back to either a projection with no disclosed underlying data for the target year, or a stale pre-AI-Overviews study (2019-2020) being re-cited with a new year attached. No primary source publishes a dated, sourced zero-click percentage for any past or current month at the precision these claims imply — dropped entirely.
- A specific percentage of ChatGPT/AI-chatbot answers that cite or recommend local businesses correctly vs. hallucinate details (hours, pricing, address)
- We searched for a primary, methodology-disclosed study measuring AI chatbot accuracy specifically for local-business facts (hours, address, pricing). We found only vendor blog posts (GEO/AEO tool companies) running their own uncontrolled samples with no published methodology, sample size, or query set — this did not meet the bar for a citable primary source.
- Similarweb's own first-party report page and dashboard for the June 2025 AI-referral figures (as opposed to TechCrunch's summary of it)
- The specific Similarweb blog/report URL surfaced in search results (similarweb.com/blog/insights/ai-news/ai-referral-traffic-winners/) could not be independently re-verified via WebFetch in this session for the exact June 2025 1.13B/191B figures — TechCrunch's direct reporting of the Similarweb data was used instead, since it carried an unambiguous, quotable figure and date. Flagging this so the page can swap to Similarweb's own URL directly if a future check confirms the exact figures live there verbatim.
- A measured (not projected) percentage of Google searches that now show AI Overviews, as of a recent 2026 date
- Google's own documentation and earnings statements give user counts (2+ billion monthly users) and a query-volume-growth claim, but no publicly disclosed, dated percentage of all Google searches that trigger an AI Overview after March 2025 (when Pew's 18% figure was measured). We did not fabricate an updated percentage by extrapolating from user-count growth.
- OpenAI's own published count of ChatGPT search referral clicks or website visits driven by ChatGPT search results
- We checked OpenAI's developer docs and help center; OpenAI publishes crawler behavior and opt-out mechanics (GPTBot, OAI-SearchBot) but no referral-traffic or click-count figures of its own. Any such number found elsewhere traced back to Similarweb/Cloudflare third-party measurement, already captured above, not to OpenAI itself.
Methodology
Sources were restricted to primary documents opened directly with WebFetch in this session: Pew Research Center's own July 2025 short-read (browser-instrumented panel data, not a survey), Google's, OpenAI's, Anthropic's and Perplexity's own current crawler/robots.txt documentation, two Cloudflare Radar blog posts with disclosed methodology and measurement windows, Google's own Q2 2025 earnings-call blog post, the Reuters Institute's 2026 Digital News Report executive summary, and TechCrunch's direct reporting of a Similarweb dataset (used because the specific Similarweb report page could not be independently re-verified in this session — flagged in dropped).
Rows were selected to cover three distinct claims that get conflated in this topic: (1) what happens to a click once an AI answer appears (Pew), (2) how AI platforms actually treat a website's content at the crawler level relative to what they send back (Cloudflare, platform docs), and (3) the platforms' own stated scale (Google, Similarweb-via-TechCrunch). Where a figure specifically describes news publishers (Reuters Institute) or the top 1,000 global websites (Similarweb) rather than local service businesses, that is called out explicitly in the first sentence of the row's note rather than silently generalized.
Every zero-click-by-a-future-year forecast, every vendor's own uncontrolled 'AI visibility' study, and any number that could not be traced to a page this session actually opened were excluded — see dropped. The Cloudflare crawl-to-referral ratios deserve one more caveat beyond the individual rows: they are per-network snapshots over specific windows, and the Anthropic ratio alone swings by more than an order of magnitude between a one-week sample and a six-month trend from the same publisher — that disagreement is left visible below rather than smoothed into one number.
Sources
- Pew Research Center, "Do people click on links in Google AI summaries?" 4 rows
- Cloudflare Blog, "From Googlebot to GPTBot: Who's crawling your site in 2025" 2 rows
- Google, "Alphabet Q2 earnings call: CEO's remarks" 2 rows
- Cloudflare Blog, "The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals" 1 row
- Google Search Central, "AI features and your website" 1 row
- Google for Developers, "Google's common crawlers" (Crawling infrastructure docs) 1 row
- OpenAI Developer Platform, "Overview of OpenAI Crawlers" 1 row
- Claude Help Center, "Does Anthropic crawl data from the web, and how can site owners block the crawler?" 1 row
- Perplexity Docs, "Perplexity Crawlers" 1 row
- Similarweb data, reported by TechCrunch, "AI referrals to top websites were up 357% year-over-year in June, reaching 1.13B" 1 row
- Reuters Institute for the Study of Journalism, Digital News Report 2026 — Executive Summary 1 row
How to cite this page
Husky Digital. AI search and local business: measured, not predicted. Last verified September 9, 2026. https://husky-digital.com/research/ai-search-and-local-business/
The same dataset, machine-readable, without the interface: CSV · JSON