Something Inc.Schedule a free consultation
GEO

Meta's AI crawler is building a fourth web index

A developer found Meta bots in his logs and drew the obvious conclusion. If it holds, the number of companies with an independent picture of the web goes from three to four, and your crawler policy is out of date.

JBJosh BernsteinManaging Partner · AUG 15, 2026 · 11 MIN READ
4
companies plausibly running an independent web index if this holds: Google, Microsoft, OpenAI, Meta
~9B
quarterly requests attributed to Meta crawlers in Anicca's August 11, 2026 roundup
2.99x
more unique URLs reached by Googlebot than by Meta-ExternalAgent, Cloudflare Radar, January 2026
1.70x
more unique URLs reached by Googlebot than by ClaudeBot over the same period
TL;DR · 60 SECONDSIndie developer Pieter Levels published log evidence of Meta crawlers hitting his sites and argued Meta is building its own web index so its AI does not have to route web searches through Google. Search Engine Roundtable covered it on August 10, 2026. Treat the conclusion as a strong inference rather than a confirmed fact. The planning consequence is real either way: a fourth independent crawler means a fourth independent reading of your site, and most enterprise robots.txt files were written for a world with two.

Building a web index is one of the most expensive things a technology company can decide to do, which is why almost nobody does it. Google has one. Microsoft has one. OpenAI built one rather than keep renting Bing. If the log evidence that surfaced this week reads the way it appears to, the Meta AI crawler is the front end of a fourth, and that changes the arithmetic of who gets to see your site directly.

The evidence is thin but specific. Pieter Levels, an independent developer who runs a portfolio of small sites, posted screenshots of Meta crawler activity across them and put forward the reasoning himself: Meta wants its own web index so that when its AI performs a web search, the query does not end up at Google, where Google could then use it for its own training. Barry Schwartz covered it at Search Engine Roundtable on August 10, noting the history of Meta's earlier and unsuccessful search partnership with Bing.

What turned up in the logs

One developer's server logs are not a confirmation, and it is worth being precise about what is established. Meta crawlers hitting a set of sites is observed. Meta building a general web index for AI retrieval is an inference from that observation plus a plausible motive. Meta has not announced anything. If you repeat this in a client meeting, repeat it with that hedge attached.

That said, it is a good inference. The motive Levels describes is the same one that pushed OpenAI off Bing: an AI company that routes its retrieval through a competitor's index hands that competitor a live feed of what its users are asking. For a company that has spent two years building frontier models and has no search surface of its own, owning the retrieval layer is the missing piece.

CLAIMSTATUSSOURCE
Meta crawlers active across independent sitesObserved, screenshots publishedPieter Levels, via Search Engine Roundtable, Aug 10, 2026
Motive is avoiding Google as a retrieval intermediaryLevels' stated reasoning, not a Meta statementSame coverage
Meta is building a general web indexInference, not confirmed by MetaNo primary source
Meta crawlers at roughly nine billion requests quarterlySecondary report, not independently verified hereAnicca weekly roundup, Aug 11, 2026
Googlebot reached 2.99x more unique URLs than Meta-ExternalAgentPublished measurementCloudflare Radar, January 2026

Why the Meta AI crawler matters more than the headline

The competitive story is the least interesting part. What matters operationally is that each independent index produces an independent reading of your site, and those readings already disagree more than most teams realize. We have documented engines disagreeing on which sources to cite for the same query, and index divergence is the mechanism underneath it. Two engines that crawl on different schedules, render JavaScript differently, and apply different canonical logic will end up believing different things about the same page.

Add a fourth and the divergence widens. Practically, that means a site fix does not propagate uniformly: you correct a product page, Google reflects it in days, one assistant reflects it in weeks, and one is still quoting the old claim in three months. Anyone measuring AI visibility as a single number is averaging over four different pictures of reality.

THE PLANNING CONSEQUENCECoverage planning has to be per-engine now, and it has been trending that way for a year. The per-engine framework we use for generative engine optimization engagements exists because a single global mention rate hides exactly this: strong presence in two engines and total absence in a third averages out to a number that looks acceptable and describes nothing.

A fourth index changes the robots.txt calculation

Most enterprise robots.txt files fall into one of three states: written in 2019 and never touched, hardened against AI crawlers in a panic during 2025, or maintained by someone who reads release notes. Only the third group is in good shape, and a new crawler arriving is the event that exposes the other two.

The default posture in a wildcard-heavy file is accidental. A blanket disallow written to stop training scrapers will also stop retrieval crawlers, which is the distinction we drew in training crawlers versus retrieval crawlers. Blocking a training bot costs you nothing you can measure. Blocking a retrieval bot removes you from an answer surface. Those two decisions live one line apart in the same file and get made as if they were the same decision.

1Enumerate before you decidePull ninety days of logs and list every distinct bot user agent by request volume and unique URLs touched. Most teams find at least one crawler they have never heard of and at least one they thought they had blocked years ago.
2Split the file by purposeWrite explicit rules per user agent rather than relying on wildcards. A new crawler should hit a default you chose deliberately, not a rule written for a different bot in a different year.
3Re-run it quarterlyNew named agents appear faster than robots.txt files get reviewed. Quarterly is the slowest cadence that keeps a policy current, and a calendar reminder is the entire implementation.
CONFIG
The wildcard trapA rule written as a blanket disallow for anything matching a bot pattern will catch crawlers that did not exist when it was written. That is the intended behavior of a wildcard and the unintended behavior of a policy, and the two only diverge when something new shows up.
EDGE
The CDN layer nobody checksRobots.txt is advisory. Your edge rules are not. Several teams we audit have a permissive robots.txt and an edge configuration that returns a challenge page to anything without a browser fingerprint, which blocks well-behaved crawlers that were reading the file correctly.
OPS
Rate limits read as blocksA crawler that gets throttled hard enough stops trying. From the outside that is indistinguishable from a disallow, and it will not appear anywhere in your robots.txt review because the decision lives in an infrastructure config owned by a different team.
HYGIENE
The staging leakEvery new crawler discovers the same forgotten staging subdomain, and now four indexes hold a copy of your unfinished positioning instead of two. Enumerate what is publicly reachable before you worry about what is being crawled.

Meta AI crawler volume against the referral you get back

The economics are the honest objection to letting any new crawler in, and they are getting worse rather than better. Anicca's August 11 roundup attributed roughly nine billion requests quarterly to Meta crawlers with minimal referral traffic in return. Treat that figure as a secondary report rather than a measurement you can defend, but the direction is consistent with everything else in this category.

Googlebot100%
ClaudeBot59%
Meta-ExternalAgent33%

Relative unique-URL reach by crawler, expressed as a share of Googlebot's reach, derived from the Cloudflare Radar comparison published for January 2026. Shown to compare coverage breadth, not request volume.

Set that against what we found in the crawl-to-referral economics work: the ratio between pages crawled and visitors returned runs into the thousands to one for several AI companies, and in the worst measured case into the tens of thousands. On a pure bandwidth-for-traffic trade, none of these crawlers pay their way. That is why publisher-side metering exists and why it keeps growing.

But bandwidth-for-traffic is the wrong ledger for most B2B companies. If a buyer asks an assistant which vendors handle a problem and your competitor is named because your pages were readable and yours were not, the cost of the block is a deal, not a server bill. Publishers with millions of pages and ad-funded models face a genuinely different calculation from a software company with four hundred pages and six-figure contracts. Most crawler advice on the internet is written for the first group and read by the second.

SITE TYPECRAWL COST MATTERSCORRECT DEFAULTTHE REAL RISK
Large ad-funded publisherYes, materiallyMeter or negotiate accessGiving away the archive for nothing
B2B software, few hundred pagesNo, negligibleAllow retrieval bots explicitlyAbsence from answers where buyers decide
Ecommerce catalogSometimes, at scaleAllow, and feed structured product dataCompetitor specs quoted instead of yours
Docs and developer contentNoAllow everything that retrievesBeing wrong in an assistant's answer forever

What to do before the index solidifies

There is a version of this article that ends with a recommendation to block Meta until it explains itself, and it is the wrong recommendation for almost every company that will read this. Blocking is a lever with an asymmetric payoff: the saving is a bandwidth line you were never going to notice, and the cost is being absent from a retrieval surface that may matter in eighteen months. Reserve it for cases where the content itself is the product and the archive is the asset. Everywhere else, the correct posture is deliberate permission with good instrumentation, so you can see what a crawler took and what it sent back.

New indexes are most malleable early. The first crawl of your site establishes a baseline understanding that later crawls update incrementally, and correcting a wrong early impression is materially harder than establishing a right one. This is the same reason a poorly handled site migration haunts a domain for a year.

So the window is now, and the work is unglamorous. Make sure the pages that define what you do are reachable without JavaScript, state your category and differentiators in plain sentences near the top rather than only in a hero graphic, and keep your entity markup consistent across the site so a new crawler resolves you to one company rather than three. We ran exactly this sequence for Zenity ahead of a category that was still being named, and the point of doing it early was that there was nothing to correct later.

For companies in fast-defining categories, the stakes are higher than for established ones, because a new index forms its idea of a young category from whatever it finds first. That is the argument we make to B2B software teams who think GEO is a next-year problem: next year the index will already have an opinion.

DO THIS NEXTThis week: pull ninety days of bot logs, enumerate every user agent by volume and unique URLs, and rewrite robots.txt with explicit per-agent rules instead of wildcards. This month: confirm your category-defining pages render their core claims without JavaScript and carry consistent entity markup. This quarter: put a recurring calendar item on crawler policy review, because the list of names will be longer in ninety days than it is today.

One developer's logs are a weak basis for a strategy and a perfectly good basis for a check. Spend the afternoon on the crawler enumeration whether or not Meta is doing what it appears to be doing. The original coverage is at Search Engine Roundtable's August 10 report.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

JB
Josh BernsteinMANAGING PARTNER, SOMETHING INC.

Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.

Free consultation

Let us be the last SEO agency you ever work with

A 30 minute call and a free audit of your SEO and GEO position. You keep the findings either way.