Do Listicles Win AI Citations? What the Evertune Data Shows

Say a statistic often enough and people start to treat it as a plan. A figure about listicles and AI citations now circulates in vendor guides, and content leads are asking whether their whole library should be rewritten as ranked lists. We followed the numbers back to their pages. What follows records what we found for each, and which ones survive the trip.

The piece is aimed at content leads weighing that rewrite. Our reading covered an AirOps guide, an Evertune post and a Peec study, and we ran no experiment of our own. All three companies sell something connected to AI visibility, and so do we, as a content software vendor, so read every conclusion with that in mind.

A quick tally before the detail

Of five claims, two are unverified, two hold only partly and one holds as a median. The 21.9% share of citations and the 63% share of model citations, both from the AirOps guide, could not be traced to a primary page. The "about half" finding for ChatGPT and the idea that self-ranking lists are a safe shortcut hold only in part. The 941-word figure from Evertune holds as a median, though the range around it is wide.

The 21.9% figure

An AirOps guide published on 20 June 2026 states that listicles account for 21.9% of all AI citations, crediting a May 2026 Evertune study of 25,000 cited URLs as relayed by Search Engine Land. The guide gives no sampling method or classification rules. Evertune's own post, from 28 April 2026, does not contain the number, and a search for the 25,000-URL study turned up only a secondary write-up with different figures, so the number may belong to another cut of the data.

That makes it unverified. We do not say it is wrong; we say it cannot be traced, and an untraceable figure does not belong in a deck.

The "about half" finding

This one has a primary source. Evertune's April post examined more than 40,000 URLs that ChatGPT cited heavily on its platform over 60 days, across nearly 50 product categories, and about half of the URLs were listicles; 58% of the listicles were ranked lists of the "10 Best CRM Tools" kind. Statista, Forbes and Wikipedia were the most cited list domains. The post leaves "heavily" undefined, covers only ChatGPT and reports correlation.

The finding holds in part. It is a real analysis of a stated sample, but the sample is one vendor's, one model's and product-focused, so it says little about AI citations as a whole. Anyone selling in a category full of comparison queries has reason to test a fair comparison page. Nobody else has reason to generalise.

The 63% figure

The same AirOps guide puts listicles at 63% of citations across ChatGPT, Gemini and Perplexity, using a different base from the 21.9%. We found no primary page with this number, and a secondary article that mentioned it could not be reconciled with Evertune. Treat it as a lead for your own test, and keep it out of anything you plan to defend.

The 941-word figure

Evertune also reports page-structure medians for cited pages, trimmed to the middle half after outliers are removed. The median was 941 words, but the middle half spanned 127 to 1,818 words, with four H2s and fifteen external links at the median. As a median for that sample the number holds. The width of the range is the more useful message, because it shows that cited pages come in many lengths. Padding to a target would be a mistake.

The self-ranking shortcut

If lists get cited, why not publish one that puts your own product first? Peec examined this in software reviews between December 2025 and February 2026. Across 232,000 citations from 13,000 listicles, about 11% came from self-promotional lists: 3.6% on ChatGPT and roughly 10% on Google AI Mode and Perplexity, and the share stayed flat for twelve weeks. So the tactic gets cited and shows no sign of being corrected. Peec still advises against it, pointing to reputational risk, and recommends earning places in third-party lists.

That is a partial hold. The study covers one niche, and its trend claim rests on a summary, not weekly data. Being cited and being wise are different questions.

What the five share

Each number lost detail on its way from source to summary. URLs examined became citations, a single model became "AI", a median became a target, and a tactic that gets cited became one that is safe. The figures may not be wrong. They have simply lost their method.

Our call: rank rivals honestly or do not rank them

The data is easy to read as a recipe, and the self-ranking list is the laziest recipe. We would build comparison pages that treat competitors fairly, say in the first paragraph that you have an interest, and explain who each option suits, including when you are the wrong choice. The price is that you will sometimes send a buyer elsewhere and lose a guaranteed top slot.

We think that is a fair trade. Readers tend to discount a list that ranks its own author first, and a list that misleads them carries the reputational risk Peec raises. We would revisit this if long-run data showed self-ranking pages converting buyers with no damage to trust, which neither dataset measures. Meanwhile, build one fair page in one category and watch it against a stable prompt set, following the measurement guide.

How to build the fair version

A comparison page worth citing has a few plain features. It opens by saying who wrote it and what interest they have. It names the criteria it uses before it ranks anything, so a reader can disagree with the weighting. It includes real competitors, including the one buyers most often consider, and says what each does well. It states who should not choose you. It shows a date and says what changed since the last update. And it links to the sources for any claim about a rival, because a statement about a competitor that you cannot support is both a trust problem and a fairness problem.

None of this guarantees a citation. It does make the page one that a sceptical buyer would forward to a colleague, which is a better test than whether a model quotes it. Review the page whenever a rival changes pricing or features, since a stale comparison is worse than none.

What would settle it

A published, reproducible method would resolve most of this: the prompt set, model versions, dates and counting rule, so anyone could rerun the work. Until then, treat the figures as a hypothesis. Structured, specific, comparison-style pages appear often among heavily cited ChatGPT URLs in product categories. That is worth a test and nowhere near a law of AI search.

A five-step routine for any statistic

Find the page the number is credited to. Search it for the exact figure. Read the method, including sample, model, period and who chose the data. Ask who profits from the finding. Then restate the claim in your own words with its limits, or drop it. The same routine runs through our earlier GEO claim check, and the GEO guide shows where this question fits in the wider plan.