The Evergreen Retreat Is a Mistake

There is a logic to the retreat. If an AI Overview can answer “how do I reset a router” in a paragraph, why pay a writer to answer it in eight hundred words? If the explainer no longer earns a click, cut the explainer and put the money into subscriptions, events, and the newsroom. The Reuters Institute’s 2026 survey of media leaders documents exactly this move, and Press Gazette’s reporting on individual publishers documents the layoffs that came with it.

We think the logic mistakes a subset for the whole. The evergreen content that AI replaces is evergreen content that was already thin — a paragraph of consensus stretched to a page. The evergreen content that AI cites is the opposite: dense with specific, sourced, current facts that an answer cannot produce on its own. Cutting the category loses the second to get rid of the first. This report lays out the data behind both halves of that claim.

1. The retreat, measured

The Reuters Institute asked 280 senior media leaders in 51 countries what they planned to do more and less of in 2026. On a net basis — the share planning to increase minus the share planning to decrease — service journalism scored −42, general news −38, and evergreen content −32. The same leaders named paid content (76%), native and brand advertising (64%), and events (54%) as primary revenue priorities.

Source: Reuters Institute, Journalism, Media, and Technology Trends and Predictions 2026; net score = share planning to increase minus share planning to decrease, n=280.

Press Gazette’s February 2026 survey of U.K. publishers put faces on the numbers. Service journalism — described by one executive as “factual, non-debatable” content such as car specifications and product-recall information — had taken the greatest losses. One specialist publisher reported severe declines on evergreen and explainer content and cut staff in response. Others reported much smaller overall effects, in part because branded search drives more than half their visits.

2. What AI actually replaces

The categories publishers are cutting are the categories where AI answers appear most — and the data on why they appear there is the key to what survives.

Ahrefs’ analysis of 146 million results pages found an AI Overview on 47.3% of definition queries, 57.9% of question queries, and 59.8% of “why” queries — against 6.3% of news queries. By intent, 99.9% of Overview-triggering keywords were informational; by topic, science (43.6%) and health (43.0%) led. Pew’s panel found 60% of question-format searches produced an AI summary and 8% of one- or two-word searches did.

Source: Ahrefs, What Triggers AI Overviews? (Law, November 10, 2025); 146 million SERPs.

A definition, a “why,” a how-to: these are questions whose answers are general knowledge, already present in the model’s training and in a thousand near-identical pages. The assistant does not need a source to answer them well, and when it names one, the click rarely follows — Pew measured 8% of searches with an AI summary producing a click, against 15% without. That is the content the retreat is right about. Its economic value has gone, and it is not coming back.

But look at what the same systems do when the question needs a specific fact. The Ahrefs citation data shows 31% of AI Overview citations coming from pages outside the top 100 — retrieved for the fact they hold, not their rank. Government sources appear in AI summaries at three times their share of ordinary results (Pew). Cited pages average 25.7% fresher than organic results (Ahrefs, 16.975 million citations). And in the peer-reviewed GEO benchmark, adding statistics to a page raised its visibility in generated answers by 33%, adding citations by 28% — while keyword stuffing cut it by 9%.

The same query — “how good is this school’s nursing program” — can be answered from general knowledge or from data. The general-knowledge answer needs no source. The data answer needs the page that holds the graduation rate, the cohort comparison, the ranking, and the date. Only one kind of evergreen is in that page.

3. Thin sites lost most

Chartbeat’s data, reported by Axios in March 2026, is the clearest evidence that depth is what is being rewarded. Over two years, search referral traffic fell 60% for small publishers (1,000–10,000 daily pageviews), 47% for mid-sized ones (10,000–100,000), and 22% for large ones (over 100,000). Across the whole network, weekly pageviews fell only about 6% — the traffic did not vanish; it moved.

Source: Chartbeat data reported by Axios, March 17, 2026; thousands of mostly news and media sites.

Similarweb data reported by Press Gazette points the same way at the category level: the “newspapers” category was down 11% year on year while “world news and media” was up 4%, and organic traffic to the top 40,000 U.S. sites overall was down only 2.5% in 2025. Winners exist. They are the sites with the authority and the depth to be the source when the answer needs one.

4. Summarizable versus citable

The practical question for an editor is not “is this evergreen?” but “could an assistant produce this page without us?” That question sorts evergreen content into two kinds.

	Summarizable evergreen	Citable evergreen
What it is	Consensus knowledge restated: definitions, generic how-tos, explainers, listicles without data	Specific facts about specific entities: statistics, comparisons, rankings, records, specifications with sources and dates
Can a model produce it alone?	Yes — it is in the training data a thousand times	No — the fact exists only where it was computed or recorded
What an AI answer does with it	Paraphrases it; rarely links	Retrieves it; must attribute it
What happened to its traffic	Collapsed with AI Overviews	Redistributed toward the deepest, freshest source
Cost to maintain by hand	Low, but pointless	High — which is why it is usually thin
Cost to maintain from data	—	Near zero per page once templated; refreshes with the data

The middle row is the retreat’s error in one line. “Factual, non-debatable” content is not what AI replaces. Generic factual content is. Car specifications for one model — trim levels, dimensions, towing capacity, dated and sourced — are exactly the kind of fact an answer must retrieve from somewhere. The reason they stopped earning traffic is not that the facts became worthless; it is that a paragraph-long page holding four of them was not the best source, and the assistant found a denser one.

5. What to build instead of cutting

The evergreen that survives has four properties, all visible in the citation data: it is about a defined entity, it holds specific and sourced figures, it is current, and it belongs to a body of coverage rather than standing alone. Those are expensive to produce by hand at any scale, which is why most evergreen content never had them and why cutting it feels costless.

They are cheap to produce from data. A knowledge graph of public and licensed sources, rendered through templates that writers author and review, produces a page per entity with every figure sourced and dated, regenerated when the data changes, and linked into the cluster around it. The economics invert: the marginal page costs almost nothing, the maintenance is automatic, and the output is the dense, fresh, citable kind. That is how a single flagship site came to publish thousands of distinct rankings and a ten-site partner network 1.8 million ranking lists, from a team that shrank while its output grew tenfold.

For a publisher with authority to spend — a strong brand, a large archive, a newsroom that already wins the headlines — the evergreen layer is the highest-return investment left in organic search, precisely because everyone else is retreating from it. Our report on news and evergreen shows the mechanism: headlines drive links and attention to permanent pages; permanent pages accumulate the authority that lifts the next headline.

Methodology and caveats

Reuters Institute figures are from a non-probability survey of 280 senior media leaders in 51 countries (November–December 2025); net scores are share increasing minus share decreasing. Press Gazette publisher accounts are from its February 2026 reporting (updated July 2026) and are self-reported by the publishers quoted. Ahrefs AI Overview trigger rates are from 146 million SERPs (November 2025); citation-rank overlap from 863,000 keywords (March 2026); freshness from 16.975 million citations (July 2025). Pew figures are from a March 2025 behavioral panel of 900 U.S. adults. Chartbeat size-tier figures were provided to Axios (March 2026) and cover thousands of mostly news and media sites; Similarweb category figures are as reported by Press Gazette via Graphite. GEO benchmark figures are position-adjusted word count in generated answers (KDD 2024). Our figures on ranking lists and team productivity are from internal reporting compiled in 2022.

What to do about it

1. Sort your evergreen before you cut it. For each page type, ask whether a model could write it without your data. Cut what it could. Keep — and deepen — what it could not.
2. Replace the generic explainer with the entity page. One page per thing your audience decides about, holding every specific figure with a source and a date. That is the page an answer has to cite.
3. Add the comparison. A fact against its cohort — better or worse than peers, up or down from last year — is the insight assistants cannot generate and readers actually want.
4. Refresh on the data’s schedule, not the calendar. Freshness is measurable and rewarded; automate it.
5. Build clusters, not islands. Link the entity pages to the rankings and comparisons above them and the specifics beneath. Depth across a subject is what separated the 22% decline from the 60%.
6. Spend the retreat’s budget on depth. The money leaving service journalism is the money that funds the data layer — and the field is emptying out at the moment the citation data says depth wins.

Sources

1. Reuters Institute — Journalism, Media, and Technology Trends and Predictions 2026 (Newman, January 2026) — Net content-strategy scores; revenue priorities; n=280
2. Press Gazette — Publishers say Google search traffic in 'managed decline', not dead (Tobitt, February 5, 2026; updated July 6, 2026) — Service journalism hardest hit; Similarweb category figures via Graphite
3. Ahrefs — What Triggers AI Overviews? (Law, November 10, 2025) — 146 million SERPs
4. Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results (July 22, 2025)
5. Ahrefs — Google AI Overview citations and top-10 rankings (Linehan, March 2, 2026)
6. Ahrefs — Do AI assistants prefer to cite fresh content? (Law, July 28, 2025)
7. Aggarwal et al. — GEO: Generative Engine Optimization (KDD 2024)
8. Search Engine Journal — Search Referral Traffic Down 60% For Small Publishers, Data Shows (Chartbeat data via Axios, March 2026)
9. Media Factual — internal reporting on ranking lists and team productivity (compiled 2022)