Blog · By industry

AEO for publishers losing clicks

Published 5 June 2026 · Updated 12 October 2026 · 8 min read

AEO for publishers losing clicks
Photograph: Bank Phrom / unsplash

The short answer: A publisher losing clicks to generated answers has two separate problems, and conflating them wastes a quarter. One is structural: the answer satisfies the reader, so the click never happens, and no markup fixes that. The other is recoverable: your pages are reachable, quotable and attributed, or they are not. AEO for a publisher means winning the attribution while you rebuild the business model around something other than the click.

Start with the size of the structural part, because most newsroom panic is aimed at the wrong half. Pew Research Center tracked the real browsing of 900 US adults across 68,879 Google searches in March 2025. On searches that produced an AI summary, a traditional result link was clicked on 8 percent of visits. On searches without one, 15 percent. That is close to half the clicks gone on summarised queries, and it is not a penalty you can appeal.

The loss splits into three buckets, and only two are yours

Satisfied demand. The reader asked what time the match starts, or what a tariff rule means, and the answer was two sentences long. Those clicks were never going to survive a surface that can answer in place. Chasing them is the main way publishers waste money in this cycle.

Displaced attribution. Your reporting is in the answer. Your name is not, or it sits below three aggregators who rewrote your story faster than you updated it. This one is winnable and it is where the work pays.

Mechanical exclusion. You are not in the answer because a bot rule, a snippet directive, a render path or an origin timeout kept you out. Boring, invisible in analytics, and usually fixable inside a sprint.

A newsroom that cannot say which bucket a given decline sits in will buy the wrong remedy. The split differs by vertical too, which is the whole argument of AEO by industry: where the work differs.

The snippet directive most publishers set years ago

Here is the mechanism that catches publishers more than any other sector, because publishers are the people who went to war over snippets.

Google's documentation on AI features states that to appear as a supporting link in an AI Overview or AI Mode, a page must be indexed and eligible to be shown in Search with a snippet. Google documents no second eligibility path for those two surfaces. Which means the preview controls you already use, nosnippet, data-nosnippet and max-snippet, do double duty on Google's AI surfaces. Other engines publish their own rules and this particular coupling is Google's, not a shared standard. A max-snippet:0 left over from a 2019 European licensing fight, or a data-nosnippet wrapper around the lede that someone added to protect a paywall funnel, quietly removes the passage an engine would have lifted and credited.

Go and grep for it. Check the robots meta tag on your article template, the X-Robots-Tag header your CDN adds, and any data-nosnippet attributes in the article body partial. These directives tend to be inherited from old legal positions, and often nobody on the current team owns them. We have found stale snippet caps still shipping on publisher article templates during our own audits, where the fix was one template line. That is a handful of sites we happened to look at, not a survey, so treat it as a thing worth checking rather than a claim about how common it is.

The honest counterpoint: a wider snippet does hand more of your text to the answer. That is a real trade. Make it deliberately, per section, rather than discovering that a global cap set by a lawyer six years ago is why the archive never gets cited.

Crawl volume is not readership, and the ratio is the number to watch

Publishers tend to look at referral traffic from assistants and conclude the surface is worthless. The more useful figure is the ratio of crawls to referrals, because it tells you what you are trading.

Cloudflare published this measurement across its own network for January to July 2025. In July, Google's crawlers made roughly 5.4 requests for every referral sent back. OpenAI's ran at about 1,091 to one. Anthropic's, after an 87 percent fall in that ratio over the six months, still sat near 38,000 to one. Those are aggregates over the sites Cloudflare fronts, not a reading of your site, and your own numbers will differ with your traffic mix. The spread is still the useful part: the exchange rate differs by orders of magnitude depending on which system is reading you, and it moved enough inside seven months that a decision made on last year's figure is already stale.

You can compute your own version this week. Split your access logs by user agent, count requests per bot per day, then count referrals arriving from each assistant's domain. Two columns, one ratio. That number, per engine, is what should drive any blocking decision, not a board-level feeling about scraping.

Separate the bots properly while you are in there. Perplexity documents two distinct agents: PerplexityBot, which builds its search index, and Perplexity-User, which fetches a page because a human just asked something that needs it. They publish IP ranges as JSON for both. Inside Perplexity, a Perplexity-User hit is the closest thing to a reader that the surface produces, and if your WAF treats it the same as an index crawler you are turning away live demand and counting it as a win. The equivalent split for other engines has to be read off each engine's own crawler documentation, since the user agent alone does not tell you whether a human is waiting at the other end.

Freshness is your advantage, and your archive is your asset

Two different games, run by two different desks.

Breaking coverage competes on being the version that exists when the question is asked. Speed of publication, a clean timestamp, a dateModified that actually changes when you update, and an update log inside the article all matter more than any keyword decision. The wire copy you file at 09:12 is the citation candidate at 09:14.

Explainers, reference pages and the archive compete on being the clearest statement of a durable fact. They stay retrievable long after publication, which is why they matter beyond the traffic they show. They also rot. A 2021 explainer with 2021 numbers and no visible update date is the page most at risk of being passed over for someone's fresher rewrite of your own reporting. We can see that outcome in the answers we sample, not the ranking logic behind it. Regulated verticals feel this hardest, which is why AEO for fintech reads as an argument about proof and dates rather than about content volume.

One more risk specific to publishers: the redesign. Media sites replatform constantly, and a URL change severs the link between the archive and everything that ever cited it. Read migrating a site without losing citations before, not after. The pattern of who gets credited in this market is set out in news sites, publishers and the citation economy.

What to measure instead of sessions

Pick forty questions your desks genuinely own. Run each several times, across more than one engine, and record three things: whether you were named, whether the claim attached to your name was accurate, and which source got the citation when you did not.

That third column is the one that changes editorial behaviour. When an aggregator who paraphrased your scoop is cited instead of you, the cause is usually structural rather than mysterious: they published a cleaner standalone statement of the fact, or their page loaded faster, or yours sat behind an interstitial. The self-check walks through ten of these tests with no tooling.

Honest limits

What this post does not establish. The snippet mechanism above is documented by Google for AI Overviews and AI Mode only, and we have not verified that any other engine couples retrieval eligibility to snippet settings the same way. The crawl-to-referral figures are Cloudflare's network aggregates for January to July 2025, not a measurement of your site, and the Pew figures are one panel of 900 US adults in March 2025 rather than the whole web. Our own observation that stale snippet caps still ship on publisher templates comes from the audits we happened to run, with no sampling frame behind it, so it says nothing about how widespread the problem is. We also cannot measure what a citation is worth to a publisher, because these surfaces send almost no referrer data, and we cannot see how private licensing arrangements change retrieval. Everything above should be re-run, not set once, since the engines revise retrieval behaviour without notice.

FAQ

Should a publisher block AI crawlers? Only after computing your own crawl-to-referral ratio per agent, and only agent by agent. A blanket block can remove you from answers where your competitors still appear, and it does not stop content reaching models through third parties who quoted you.

Do we lose less if we widen snippets? You hand over more text and you become more quotable. Pages whose job is to build authority usually win from a wider snippet; pages whose job is conversion often do not. Decide per template.

Is the click loss permanent? The portion where a two-sentence answer was enough looks permanent. The portion caused by exclusion or weak attribution is not, and it is often the larger share on explainer and archive pages.

What does an audit actually hand over? A report. It runs a fixed prompt set of about 45 prompts across five engines, with retrieval on and off, and keeps the evidence behind each finding so an editor can argue with it. The sample audit report shows the shape.

Sources

Related reading

AEO by industry: where the work differsAEO for fintechMigrating a site without losing citationsNews sites, publishers and the citation economyAEO when every claim needs approvalAEO for agencies, adding it to a retainer