AEO for media publishers: what to make answerable, what to protect, and how to be the cited source

AEO for media publishers is answer engine optimisation adapted to a business whose content is the product. For most organisations an answer that quotes their page is a route to the sale. For a publisher the same answer can replace the visit the page was written to earn, so every extractable passage is a commercial choice.

The advice in our other guides assumes the opposite economics. A software company or a clinic loses little when an assistant summarises its explainer, because the explainer existed to build trust ahead of an enquiry. A publisher earns from attention on the page itself: advertising, subscription prompts, newsletter sign-ups. An answer that resolves the reader's question upstream removes the moment in which any of those could happen, and a citation link does not put that moment back.

That does not make answer engine optimisation for publishers a purely defensive exercise. Some content will be answered whether or not the publisher cooperates, and being the named source is better than being paraphrased anonymously. The work is sorting the two kinds of content, then applying the right controls to each rather than the single sitewide stance most publishers inherit by default.

Publisher content sorts by what an answer does to it. Quick facts, definitions, dates, results and schedules get answered by engines regardless, so the aim there is to be the cited source. Analysis, investigation and service journalism lose most of their value when summarised, so they are worth protecting or shaping into a reason to click.

The test is simple to state and slow to apply: if an assistant gave the reader the gist of a page in three sentences, would the reader still need to visit? For a match result or the definition of a policy term, no, and the answer will be written from somebody's page. For a long investigation, a buying guide built on hands-on testing, or an analysis whose value lies in its reasoning, a three-sentence gist is a loss, and the page deserves different treatment.

Most publications contain both kinds, often on the same page, where a factual summary sits directly above the analysis that gives it value. The useful output is a content map by template and section, agreed between editorial and commercial leads, that assigns each type a default treatment. Exceptions can be handled story by story; defaults cannot, and a newsroom without them ends up making the decision by accident.

Google's snippet controls are the main lever a publisher has over AI Overviews. Google documents that nosnippet, max-snippet and data-nosnippet apply to its AI features in Search as well as to ordinary results, so they limit how much of a page those features can show. They do not remove the page from Search.

The three work at different grains. The nosnippet robots meta tag withholds text from snippets for the whole page. The max-snippet directive caps the number of characters that may be shown. The data-nosnippet attribute marks particular elements, such as the analysis beneath a news summary, so the rest of the page can still be quoted. That last one is the precise tool for the sorting described above: let the factual top of a page be quoted and hold back the reasoning below it.

The trade-off is that the same limits apply to the ordinary blue-link snippet, so a protected page shows less in normal results too, which can cost clicks of its own. There is no documented switch that removes a page from AI Overviews while leaving its Search presence untouched. Google-Extended does not do it: that token governs whether content is used for Gemini models and their grounding, while AI Overviews draw on Googlebot's crawl. Treat all of this as Google's current documentation, which changes, and not as a guarantee of how any engine behaves.

Crawler policy for publishers is a commercial decision taken per AI agent, not a single allow or block. Training crawlers, search crawlers and user-triggered fetchers do different things with a page, and restricting training while allowing retrieval, or restricting everything until a licensing deal is in place, are both legitimate positions for a publisher.

The distinction matters because the agents are separable in robots.txt. GPTBot and ClaudeBot collect material for model training. OAI-SearchBot, Claude-SearchBot and PerplexityBot index pages so an assistant can cite them in answers. ChatGPT-User, Claude-User and Perplexity-User fetch a page when a person asks about it in the moment. Blocking the training crawlers keeps an archive out of future models without removing the publisher from cited answers; blocking the search crawlers and fetchers removes it from those answers as well. Some providers document that their user-triggered fetchers treat robots.txt differently, so read each one's current documentation rather than assuming.

Licensing sits alongside robots.txt rather than replacing it. A publisher negotiating licensing terms has good reason to restrict access until the terms exist, and our piece on why llms.txt is not robots.txt sets out why publishing is the genuine exception to our general advice to allow AI crawlers. We do not re-argue it here. The point is that the decision is written down, owned commercially and enforced consistently in robots.txt and in the bot-management layer.

A publisher becomes the cited source by making each story easy to attribute and easy to trust: a named author with a real profile, accurate publication and modification dates, visible corrections, and Article or NewsArticle markup that says exactly what the page says. When several outlets report the same facts, those signals are what separate one source from the rest.

Author entities deserve more work than they usually get. A byline that links to a profile page listing the writer's beat, background and other work, connected to the Article markup by a stable identifier, lets an engine resolve who wrote the piece. Dates need the same discipline: datePublished is when the story first ran, dateModified changes when the substance changes, and neither is touched to fake freshness. A story that bumps its date for a typo fix is asserting something untrue in a machine-readable field.

Corrections are the credibility signal publishers most often hide. A silently amended page leaves the earlier version in caches and in answers already given, with nothing on the page to reconcile them. A dated correction note tells readers and engines alike which version stands, and it costs a sentence. It is also what a reader checking the record, or a regulator, expects to find.

Evergreen explainers and live topic pages are the AEO surface that keeps earning citations for a publisher, because they answer the question that persists after the news moves on. News is accurate on the day it runs and stale soon after; a topic page kept current stays the best available answer while someone maintains it.

Assistants asked about an ongoing subject need a source that states where things stand now. A single URL with a short current summary at the top, a dated timeline beneath it and links into the underlying reporting gives them one. Twenty dated news stories, each correct as of its day, give them twenty partial answers and invite the engine to assemble its own summary from someone else's page.

This is also where the commercial logic is kindest. An explainer that defines a term or sets out the background is content an engine would answer anyway, so being quoted costs little, and the citation points at a page that links onward into the reporting the publisher does want read. Keeping such pages current is editorial work with a named owner and a review date, not a one-off optimisation task.

Publishers measure AEO with two instruments: a fixed panel of prompts about the subjects they cover, run cold on each engine at a regular interval to record citation share against a set of competing outlets, and referral traffic from assistants in analytics. A screenshot of one favourable answer is a single sample and measures nothing.

The panel should follow the publication's beats, and it should include prompts for the content the publisher chose to protect as well as the content it chose to open, because the question is not only whether the publication is cited but whether its protected work is being summarised from elsewhere. Our piece How to Measure Who Gets Cited by AI sets out the method in full: cold sessions, fixed scoring rules and a fixed competitor set.

Referral traffic from assistants appears in analytics as sessions from the assistants' domains, where the referrer survives, and deserves a segment of its own. Read it against what each visit is worth to the business rather than as a volume, since the value of the visit is what the whole strategy is trying to protect. Neither instrument attributes a change to a single action, and neither should be reported as if it did.

AEO cannot fix the structural loss of clicks that answer engines impose on publishers. Answer engine optimisation for publishers can decide what is quoted, secure attribution where citation helps and protect the work that depends on a visit, but it cannot make a reader click through when the answer on the results page was enough.

We say this plainly because the opposite is often sold. No markup, snippet setting or crawler policy restores the old exchange in which a publisher supplied the answer and received the visit. What the work does is stop the losses that were avoidable: being paraphrased without credit, having protected analysis summarised because nobody set a control, and publishing a run of news stories without the explainer that would have carried the citation.

The durable answer sits outside search, in direct relationships with readers that do not depend on an intermediary deciding whether to send them. AEO is worth doing well for what it can influence, and worth being honest about for what it cannot. Citations still matter within that strategy, as one of the ways a reader first meets the publication before deciding to come back directly.