🔎 Publishers in AI answers: AI Overviews benchmarks 2026

Search results are no longer the main showcase for media: the reader gets a ready-made answer right in the interface, and the source link turns into an optional detail. Editorial teams looking for additional distribution channels benefit not only from their own traffic, but also from a platform for guest publications, where content immediately reaches someone else's audience. Below we will look at the rules generative systems use to select sources, which figures are already confirmed by research, and what an editorial team can change in the coming weeks without rewriting the entire archive.

How a publisher can get into a generative search answer

💡 Quick overview:

• Step 1: Collect a list of questions the audience asks in words, not key phrases, and check it against real suggestions in chat interfaces.
• Step 2: Convert articles into a format where the answer to the question sits in the first paragraph of a section, not hidden in the middle of the text.
• Step 3: Mark up content with FAQPage and HowTo schemas so the machine sees the structure without guessing.
• Step 4: Strengthen topical clusters: one deep library on a topic is cited more often than a dozen scattered notes.
• Step 5: Set up a separate report on brand mentions in assistant answers, without mixing it with the click report.
• Step 6: Rebuild outdated content once a quarter, because models prefer fresh dates and verifiable links.

This sequence looks boring, and that is its value. None of the steps requires a new team or a separate budget: it requires a different order of priorities within the existing editorial process. Previously an editor thought about a headline for clicks, now they think about a snippet the machine can cut out and quote in full.

What is not on the list matters too. There are no keyword density tricks and no attempts to fool the model with structured junk. Generative systems rely on signal consistency: if a site covers a narrow topic for a long time and consistently, it gets into the answer. If it jumps between topics for reach, it will be pushed out by someone who dug deeper.

What happened to the "query, answer, click" chain

The logic of classic search was linear: the user entered a query, scanned ten blue links, and picked one. The generative layer broke this chain in the middle. The answer is formed before the reader sees your headline, and sources collapse into a compact footnote block that many people simply do not expand.

The scale of the shift is clearly visible in the numbers. A Pew Research Center study based on a sample of 900 adult US users showed: when a generated summary sits above the results, an organic link is opened in 8% of visits versus 15% without it. A link inside the summary itself gets clicked in only 1% of cases. The summary itself appeared on 18% of all observed queries, meaning this is no longer an exotic scenario but an everyday one.

Post-reading behavior is even more telling. The same Pew dataset records that the session ended right on the summary page in 26% of cases versus 16% on a regular results page. At the same time, 88% of summaries relied on at least three sources. This creates the paradox media now live with: many are cited, almost no one is clicked.

This leads to a simple conclusion about metrics. If success is measured by clicks, generative search looks like a failure: according to Conductor's consolidated AEO report, AI accounts for only 1.08% of all site visits. Chasing that volume is pointless. The value lies elsewhere: the publication's name is spoken at the moment when a decision is just forming, and that effect does not fit into a sessions report.

Who gets cited and for which queries

The first thing that breaks intuition: the leader in citations and the leader in user numbers are different systems. According to Conductor benchmarks for the publishing segment, Google AI Mode accounted for 54.3% of all analyzed citations, while ChatGPT collected 19.4%, and Perplexity 13.1%. The reason is technical: AI Mode splits one question into several subqueries and pulls a separate source for each, so its average number of footnotes per answer is noticeably higher.

The second surprise concerns the relationship with organic positions. According to the same Conductor sample, the overview appeared on 41.3% of publisher queries, but 57.8% of cited pages were not in the top ten organic results. These are two different games on the same field. A page can sit on the fourth page of search and still feed the answer if it answers more precisely than the others.

The third shift is the most unpleasant for industry strategies. The competitor now comes from a different category. The model names whoever answers the question best, regardless of whether it is a publication, a retail chain, or a medical center. Planning visibility against a list of direct competitors no longer works: the list has to be built around questions, not companies.

A real case: news sites and the collapse of search traffic

The most honest example of the consequences came not from a marketing platform, but from an academic institution. The Digital News Report 2026 by the Reuters Institute at Oxford is based on a survey across 48 countries and records that weekly use of chatbots for news grew from 7% to 10% in a year. Social and video platforms as an entry point to news reached 54%. The audience moved, and it moved faster than editorial processes could adapt.

Then the unpleasant part begins. Press Gazette's breakdown of the report shows a drop in Google search traffic across a sample of more than 2,500 sites of roughly a third over the year, and up to 38% in the US market. At the same time, referrals from chatbots themselves remain below 1% of publishers' total traffic. The hole is not being filled by a new channel; it simply remains a hole.

What surviving newsrooms are doing about it is clearly visible in GIJN's piece on the report's findings. The bet is shifting to direct relationships with readers: newsletters, podcasts, apps, membership programs. Search traffic stops being the foundation and becomes a nice bonus. At the same time, the role of author recognition grows, because trust in anonymous text in a world of synthetic answers is falling fastest of all.

The lesson for any publisher, not just news publishers, is this: diversify your entry points before the chart starts going down. Rebuilding channels takes a year or more, while a drop in visibility happens in a single update.

How to measure visibility in assistant answers

The old set of metrics is useless here because it only counts those who made it through. You need a second loop: how often the brand appears in answers, for which question phrasings, with which link, and in what context. Semrush's extended visibility index is built on analysis of 126 million prompts and shows why this matters: visitors coming from generative interfaces convert about 4.4 times better than average organic traffic.

The organizational takeaway from the same Semrush study is even more stark. Among teams that merged search work and AI visibility work into a single process, 81% reported growth in traffic or leads, while among those running these areas separately, only 36% did. The difference is not in tools, but in the fact that two departments stop optimizing mutually exclusive metrics.

A practical measurement framework looks like this. First, establish a baseline list of several dozen questions critical to your topic and query assistants on them every two weeks. Second, count mentions without links separately: they don't bring traffic, but they build recognition and often precede a citation. Third, track which specific pages end up in footnotes, because that's a direct hint about which format is worth scaling.

It's also worth spelling out what not to measure. Don't try to reduce AI visibility to a single composite score: it averages out different systems with different logic and hides exactly the gaps you need to find. And don't compare numbers between quarters without rebuilding the question list, because audience phrasings change faster than your reports.

Which content formats models pick up more readily

Experience from recent months converges on several recurring patterns. The machine likes text from which it's easy to cut a self-contained piece: a short definition, a comparison table, a numbered sequence of steps, a direct answer to the question in the first sentence of a section. Anything that requires reading three paragraphs to get to the point loses.

The second pattern is verifiability. A claim with a date, a named source, and a live link gets cited more readily than a polished generalization. Here, editorial standards and the demands of generative search have aligned for the first time in a long while, which is rather good news for those who were already working carefully.

The third pattern concerns topic coverage. A single article on a narrow question has almost no chance against a library of twenty related pieces with internal linking. The model looks for confirmation across multiple points on one domain, and archive depth works as an argument in your favor.

⁉️🤔 Frequently asked questions

Should you block AI crawlers from your site to protect content?

Full blocking costs you citations, and the traffic doesn't come back: the reader will get the answer from another source. It makes more sense to split access by section, keeping explanatory materials open because they build recognition, and closing off what actually sells subscriptions. The decision is worth revisiting every few months.

Is it true that a high organic position guarantees inclusion in the summary?

No. According to publisher benchmarks, most cited pages were outside the top ten of organic results. The overlap between these two lists is partial, so optimizing for position and optimizing for citation require different work on the text, although basic technical accessibility of the page is needed in both cases.

How long does the first noticeable result take?

Usually one to two quarters, and this is tied not to the algorithm but to the reindexing cycle and signal accumulation. Updates to existing materials with a good history respond fastest; new sections without internal links respond slowest. Plan the work as cumulative, not as a one-off campaign.

Do you need to rewrite the entire archive for the new format?

No, and attempting to do so usually consumes resources without return. Pick materials that already get noticeable search traffic and rework their structure: question in the subheading, answer in the first sentence, lists and tables properly marked up. Touch the rest as facts get updated.

How do you tell visibility growth from random noise in reports?

Look at consistency across several systems at once and repeatability across different phrasings of the same question. A one-off appearance in an answer is often random, but stable citation across a group of related queries is a signal. It's useful to record not just the fact of a mention, but the specific page that ended up in the footnote.

Does an author byline affect the likelihood of being cited?

There's no public formula, but indirect data on audience trust and verifiability requirements speaks in favor of a byline. An author name with a publication history on the topic, profile links, and a bio strengthens the consistency of competence signals, and that consistency correlates noticeably with getting into answers.

Bottom line: what editorial teams should do right now

The picture is consistent. The generative layer takes the click but distributes visibility, and these two things can no longer be measured with one report. Publishers that keep counting success only in sessions will watch their charts slowly decline while their brand presence in answers remains quite alive. Publishers that set up a second measurement loop will see exactly where they're already winning.

A practical minimum for the coming month is compact: compile a list of real audience questions, rework the structure of your ten strongest materials around them, mark up FAQ and step-by-step blocks, set up regular mention tracking, and in parallel strengthen direct channels, from newsletters to partner publications on other platforms. None of these require heroics; all of them require consistency.

Start small: take one article that brings you the most search traffic and rewrite only its section headings, turning them into reader questions. Measure the result in a month. If the effect holds up, you'll have an internal argument for restructuring the entire editorial process, not just one piece of text. The other materials in this section will help you build the next steps.