Reddit is locking the door AI citations come through

On Thursday Reddit told shareholders that search traffic fell by more than a million referrals a week on average during October, and that it will test a logged-in-only version of the site in select markets outside the US this quarter. On the same day, at a Digiday event in London, People Inc.'s chief innovation officer described moving its AI partners from web crawling to a custom API because "90% of the content is junk meant for a browser to read, not an agent." One company is closing the crawl. The other is replacing it with something narrower and cheaper. Both are telling you the same thing about the assumption your GEO program is built on.

The short answer: access to the sources an answer engine reads is moving from "public crawl" to "granted permission," and permission now has terms, prices and expiry dates. That changes what you can audit. It does not change what you should publish — it changes what you should stop counting.

What Reddit's letter actually says

The headline being shared is that AI Mode killed Reddit's traffic. The filing is more careful than the coverage, and the difference matters if you are using Reddit as a proxy for how AI surfaces treat the rest of the web.

• The number: search referrals dropped by more than 1 million per week on average during October, following a third quarter in which search referral traffic was already "down meaningfully."
• The forecast: if the decline continues at that rate, total daily active uniques could fall in Q4 by a low-double-digit percentage from a base of 130 million.
• The attribution: CEO Steve Huffman said the biggest part of the working hypothesis is Google's AI Overviews and AI Mode. He also named a second factor — machine-translated content being ranked lower, mainly outside the US — and said Reddit has no visibility into whether this behaves like a normal algorithm update or a permanent one.
• The gap: neither the shareholder letter nor the earnings call splits the decline between those two causes. Anyone quoting a single number for "traffic lost to AI" is quoting something Reddit did not publish.

Then the part that belongs to your job. The login test is being described as a question about how much content people see before a prompt, including people arriving from search. It is also, in the letter's own words, a way to protect access to Reddit's corpus and adapt to emerging regulation. Huffman added that the vast majority of scraping happens through logged-out access.

Read those sentences together and it is not a UX experiment. It is a corpus-access decision wearing a UX experiment's clothes, announced by the largest referrer-dependent platform on the web, one quarter after the traffic that made it valuable started falling.

The countermove: a second door, built on purpose

While Reddit was explaining why it might close the front door, publishers were building a side entrance with a lock on it.

People Inc. went from crawling to Amazon S3 buckets — which let it push updates and retractions — and is now moving to custom APIs that return article text in Markdown plus author, published date, updated date and other provenance signals. Forbes created Markdown versions of its Vetted reviews and an "MCP layer" alongside them. The Economist is experimenting with what Digiday called a two-track internet: material outside the paywall made easier for agents to read.

Two opposite strategies, one shared arithmetic. An agent reading your HTML pays for navigation, cookie banners, related-post rails and ad slots as tokens. The publisher bears none of that cost and the reader resents all of it. So the publisher strips the page down to something an agent can afford to read, and in exchange asks for a durable relationship instead of a page view.

Cloudflare had already made that exchange literal. Its pay-per-crawl feature resurrects HTTP 402 Payment Required: a crawler either presents payment intent in headers and gets a 200, or gets a 402 with the price. Domain owners pick allow, charge, or block per crawler, and Web Bot Auth stops the crawler identity from being spoofed.

The four access modes, and what each one costs you

Most GEO advice assumes exactly one access model: pages that are public, crawlable and free. There are now four, and they fail differently.

The fourth row is the one that quietly breaks reporting. When an agent gets a fact from your API and repeats it in a sentence, there is often no link, no referrer and no session. You were the source. You cannot see it.

Two numbers this breaks in your dashboard

First: citations and clicks are different KPIs owned by different systems. Reddit is the clearest case available, because it published both halves — it is one of the most-cited domains in AI answers and it just told shareholders the click side is falling by seven figures a week. If your GEO report puts "AI visibility" and "AI sessions" in the same table, one of those rows is not measuring what the other measures.

Second: Google does not give you an AI surface to filter on. Search Console documents AI Overviews and AI Mode inside the Web search type, blended with classic results, with no filter to separate them — and every link inside an AI Overview inherits the position of the whole block, which is not a rank. In analytics, AI Mode arrivals usually carry a google.com referrer and look like ordinary organic. Assistant traffic is identifiable only from its own referrer strings.

So the honest version of "how much AI traffic do we get" is: a count of assistant referrals, a blended number for Google surfaces, and a separate citation count you collect by hand. Three systems. Anyone presenting one number is choosing which two to hide.

The access map: one page, five rows

This is the artefact I would build before touching a single piece of content. Take the five surfaces that actually cite you — not the ones you like — and record one row each:

1. Surface — the specific place the citation appears, not "Google."
2. Access mode — one of the four above.
3. Terms — who can read it, under what conditions, and when those terms were last checked.
4. Read-back — can you reproduce the citation yourself, on demand, without asking anyone?
5. Rotation — how fast the set of cited sources changes on that surface.

The rule that makes the page worth keeping: a surface you cannot read back is a blind spot, not a channel. In client work the same failure repeats — programs optimize for surfaces nobody wrote down, and the first real audit finds citations that cannot be reproduced, because the page was cited from a licensed copy, or the answer used an API, or the wall went up between the baseline and the read.

Mapping cost is hours, not months. Not mapping costs you the quarter you spend optimizing a source that stopped answering to crawlers in August.

The opinion: citation share is a rented metric

Rank, sessions and assisted conversions are yours. They can be defended with your own logs. Citation share is rented from an entity that can change its terms in a single quarter — and, as of this month, has started to.

That is why I would split the GEO budget line rather than defend one blended number. Half goes to paying for access and provenance: a clean text rendition, real author and updated-date metadata, an explicit access policy instead of an inherited one, and the discipline of writing down what each surface allows. The other half goes to owned demand — the email list, the direct visit, the returning customer — because those do not disappear when a platform decides logged-out readers are scrapers.

The uncomfortable prediction: the surfaces that cite you in 2027 will disproportionately be the ones with a contract behind them. Contracts do not come with a checklist, which means the advantage goes to whoever started mapping access this quarter, not to whoever waited for a standard.

If you want a second pair of eyes on that map, our SEO and GEO work starts with an access and citation audit before any content decision. The related mechanics are already documented here: why crawler permission and citation are separate systems, and which switch you are actually pulling when you block AI training. If your site also has to serve agents directly, the pages-and-protocols decision is where the rebuild conversation starts.

The crawl is not going away. It is becoming one lane of four — and the lanes are being priced while most teams are still arguing about word count.