AI systems read our pages 5.1 times as often as Google and Bing

The ratio

Over the last thirty days, agents belonging to AI systems made 697,475 requests to network pages against 137,570 from Googlebot and Bingbot. That is 5.1 to one. The ratio is not a curiosity about bots. It is a description of who is doing the reading that shapes what people are told.

Across the whole record the same measurement gives 1,502,750 against 177,757, or 8.5 to one. We lead on the thirty-day figure rather than that larger one, for the reason in the next section.

Two years ago a directory lived or died on Google. Today the readers that matter most arrive from Meta, OpenAI, Amazon, Anthropic and Perplexity, and they read everything, constantly.

Why we do not quote the whole-record ratio

The record opens on newly launched sites. Its first 10 days hold 781,958 requests from AI crawlers against 38,268 from Googlebot and Bingbot, a ratio of 20.4 to one. That is 52% of all AI volume on the record, concentrated in days that describe a launch rather than a rate.

Recording began part-way through the first of those days, so the period covers 9.21 logged days rather than a full 10. The daily rates below divide by that, which is why they are slightly larger than dividing the totals above by 10 would give.

What drove it is worth stating precisely, because the obvious explanation is wrong. Search crawling did not stop: across those days Googlebot and Bingbot ran at 91% of their sustained daily rate. It was the AI side that surged, to 3.7 times its sustained rate. The ratio moved because AI crawling spiked on newly published pages, not because search crawlers were absent.

A headline built on those days would therefore describe a burst of discovery rather than how these pages get read. The sustained figure is the honest one, and it is the one at the top of this page.

Every number in this section is printed above it: the totals for the period, how long the period ran, and the sustained rates they are measured against. The multiples follow from those and nothing else.

Who is doing the reading

Rolled up to the company behind the agent, the order is Meta, OpenAI, Amazon, Anthropic, Perplexity. The rollup is the comparison worth making: several of these companies run more than one crawler, so an agent-by-agent table quietly sets one company's single agent against another's whole fleet.

None of this says how any company used what it read. A fetch establishes that a page was requested. It establishes nothing else, and we do not stretch it.

Two kinds of reader

Most of the volume is index and training crawlers: Meta-ExternalAgent, Amazonbot, OAI-SearchBot, GPTBot, ClaudeBot, PerplexityBot, Bytespider, GoogleOther. They have different discovery, indexing and training purposes; a fetch does not establish how the content was used.

A smaller number are answer-time agents: ChatGPT-User, DuckAssistBot, Claude-User, Perplexity-User. These fetch a page in the seconds after a person asks a question. Across the whole record they made 9,212 requests, 7,812 of them in the last thirty days. These user-agents are associated with user-triggered requests. The log does not establish whether the page was used in an answer. That is the number the whole program is aimed at.

What this record does not cover

A site can only log a request if it serves its pages through code we run. The rest of the network publishes as static exports, and a crawler reaches those pages without ever touching anything we operate. The sites in this record carry about 171,000 of the network's roughly 480,000 published pages, about 36% of them. Our largest publication is not one of them.

So every figure on this page is a floor. The part of the network we cannot see is larger than the part we can, and the gap runs in the direction of more reading rather than less. We would rather publish the coverage and be checked on it than publish a bigger number nobody can audit.

What they read

The deepest-read pages across the network are the comparative ones: best roofers in a city, date-night venues in a suburb, makeup artists in a region. Individual business pages are read, but the lists are read first and most. Specific, traceable facts make these pages useful sources. Inclusion on a list does not ensure an assistant will recommend a business.

How to read the numbers

A request is not a citation. An agent can fetch a page and never use it. We measure citations separately, with a fixed panel, and never add the two together.

A request is not a person. Nothing in this report describes human visitors; they are measured by a different mechanism and reported separately.

The two tables below count AI crawlers only. Googlebot and Bingbot are search crawlers and are counted separately throughout. The Google row in the company table is GoogleOther and Google-Extended, which are Google’s AI crawlers rather than the one that builds its search index.