AI search statistics for home services (2026)

What we see in our own logs

Every figure above is someone else’s research. This one is ours, and it answers a question those studies don’t: do AI assistants actually come and read a small, new home-services website? Over the 37 days from 2026-07-29 to 2026-09-03, this site’s server log recorded 13 distinct crawlers fetching pages — 8 run by AI companies and 5 by search engines.

At least one AI crawler appeared on every one of those 37 days. On 34 of them, a fetch arrived with a user-agent that OpenAI and Anthropic send only when a live person has asked their assistant something — as opposed to the background crawling that builds an index.

Method: this site’s own edge log, 2026-07-29 to 2026-09-03 (UTC), transcribed 2026-09-04. The log records the first recognised crawler request of each day, so a “day seen” means that crawler fetched at least one page — not how many. 682 raw log lines became the 236 rows behind this table after dropping 4 requests from our own network, dropping 32 from 5 scanner bursts (single rented hosts on 3 networks, each claiming several different AI-crawler names within seconds — the largest 12 lines wearing 7 names in half a second, while walking snapshot pages), dropping 1 whose client IP failed the check against its operator’s published crawler range, and collapsing 409 repeated entries. Where an operator publishes such a range we check it: 131 of the 236 rows passed, and a row that failed is not in the table at all. The remaining operators publish no list, so those rows rest on the user-agent alone.

• This is one website — ours — not a survey. It shows what reached this site and generalises to no other.
• The counts are floors, not volumes: a crawler that pulled five hundred pages in a day and one that pulled a single page both show as one day.
• A user-agent is a claim. Google, Bing and OpenAI publish IP ranges we check against; Anthropic, Perplexity, Amazon, Meta, Yandex, Apple and DuckDuckGo do not, so those sightings are self-reported — including the Claude-User fetch.
• This records who fetched the pages. It is not evidence that any assistant recommended a business, which is a different question and one we measure separately.

Which page an assistant lands on when a person asks

The table above counts crawlers. This one counts pages, and only for the fetches that carry a user-agent OpenAI and Anthropic send when a live person has asked their assistant something. Over the 37 days from 2026-07-29 to 2026-09-03 there were 236 such fetches, on 34 separate days.

We publish it because it is the question a contractor should ask us before anything else: when an assistant reads a home-services site on someone's behalf, what was that person actually after? On this site, in this window, 87% landed on the front page and 7% on a city cost page — a homeowner pricing a repair, not a business owner shopping for marketing.

The last row is the one we would rather not print. We publish 47 snapshot pages naming real businesses, for the express purpose of being found by a contractor looking up their own name — and across these 37 days not one of them was opened for a live person. We keep the row at zero rather than dropping it, because we did not fail to look: that is 47 pages, 37 days of watching, and a nil return on the channel we built to reach you.

Method: the same edge log as the table above, 2026-07-29 to 2026-09-03 (UTC), transcribed 2026-09-04. 244 live-user fetches were reported and 236 counted. The 8 dropped, each for one stated reason: 2 were our own automation fetching a page to check it (our ISP, AS51896 — the exclusion is in the data, not in a footnote); 3 were a scanner on Google Cloud (AS396982) sending answer-engine user-agents from an IP outside OpenAI's published range; 3 was one fetch reported twice by a throttle that fails open, the pair 0.3 ms apart. Of the 236 counted, 233 had their client IP checked against OpenAI's published crawler range and found inside it; the other 3 are Anthropic's Claude-User, which publishes no range to check and arrived from a consumer ISP rather than our own network. A fetch whose IP is checked and found OUTSIDE the operator's range is never in this table, whatever else is true of it.

• n = 236 fetches on one website — ours. This is a server log, not a survey, and the shares describe this site's pages, not demand in any market.
• Each row is a first sighting, not a visit count. A person whose assistant read four of our pages contributes one row, and the page recorded is the first one — which is what makes it a landing page, and why every count here is a floor.
• A landing is not a referral. It records that an assistant fetched the page, not that it cited us or that the person ever arrived. Human arrivals from these fetches remain at zero.
• 37 days is a short window and the counts are small. A single busy week could reorder this table, and we will republish it when it does rather than leaving the first read standing.

What those reads led to

Both tables above count machines reading this site. This one counts people doing something afterwards, and it is the section we would leave out if we were selling you something. Over 2026-07-29 to 2026-09-03, across the 236 live-user fetches in the ledger above, every engagement counter on this site fired 0 times.

Not one form, one tool, one claim, one enquiry. The oldest of these counters has been watching since 2026-07-08 — 57 days — and has never recorded a single person. That “never” is the whole record and not just this window: every firing these counters have ever made is either in the table below or named in the exclusions under it. We publish it because the two tables above are worth nothing to you unless you can see what we do with a number that goes the other way, and because “an assistant read the page” and “a homeowner did something” are different claims that this industry routinely sells as one.

One counter is not at zero, and it is the one that measures the least. On 2026-09-02 the arrival beacon fired 5 times — it records a landing, not an action, firing once per session on the first page load. Both rows came from one search referrer to one page, on the same device from the same network, and the honest floor is one person in two sessions rather than two people. They touched nothing: every counter in the table below stayed at zero that day and has stayed there since. We are separating the two numbers rather than adding them because “somebody arrived” and “somebody did something” are the two claims this whole section exists to keep apart.

Method: the same edge log and beacon channel as the two tables above, 2026-07-29 to 2026-09-03 (UTC), transcribed 2026-09-04. The table is every counter this site can emit, not a selection — the list is generated from the type the beacon endpoint validates against, so a counter cannot be added to the site and left out of here. "Live-user reads it could have caught" is the counted landings from the table above dated on or after the day that counter shipped; for the counters older than this window that is a floor, not a total, which understates their exposure rather than ours. 571 further rows exist in the channel and are counted in no figure on this page: 569 were recorded before this site could tell its own traffic from a visitor’s — our ISP, datacentre networks and user-agent-less scrapers all still reached the counter, and no reading since has treated them as people; 2 were our own deploy smoke-test, submitting the form from a page named for it. They are named because a first-party log with an invisible filter on it is worth less than no log at all.

• This is one website — ours — and a new one. A zero here is a fact about this site’s reach, not evidence that AI-referred homeowners do not act anywhere.
• A zero is only as strong as the exposure behind it, which is why every row carries its own days and its own denominator. Five days of nothing is not eleven, and eleven is not forty-two.
• These counters see somebody who arrives and acts. They cannot see an assistant saying a business name out loud to a homeowner who then calls that business directly — that half is genuinely beyond us, and beyond anyone else selling you a number for it.
• A counter fires on a real interaction, not on a page view. Nothing on this site emits one of these links into the wild, so a crawler walking our pages cannot trip one — which is what makes the zero a measurement rather than a gap in the instrument.