The Page Shapes AI Retrieval Actually Finds: 525 Versus Pages, 17 Reviews
Of the 2,680 results ChatGPT retrieved across 60 software buying questions, comparison and "versus" pages were the single largest identifiable content type at 525 results – 19.6% of the corpus. Best-of listicles came second at 371. Pricing pages third at 278. Review pages, the format an entire industry is built on, accounted for 17 results. That is 0.6%. If you are choosing what to publish for AI search, that ranking is your brief.
What the retrieval layer actually pulled
| Retrieved page type | Results | Share of corpus |
|---|---|---|
| Other / product page | 1,161 | 43.3% |
| Comparison / versus | 525 | 19.6% |
| Listicle / best-of | 371 | 13.8% |
| Pricing / cost | 278 | 10.4% |
| Alternatives-to | 196 | 7.3% |
| Guide / definition | 132 | 4.9% |
| Review | 17 | 0.6% |
Classification was by page title, for the retrieved results whose titles could be classified. "Other / product page" is the residual: homepages, feature pages, docs, everything that is not one of the six named shapes.
Strip out the residual and the picture is stark. Among the pages that declare their purpose in the title, more than a third are comparisons or versus pages. Reviews are a rounding error.
Why versus pages win
The mechanism becomes obvious once you look at what the fan-out is doing. The model issues multiple parallel queries and receives results grouped by domain. It is not reading the page; it is reading the title and snippet, deciding whether to pull the page into context, then extracting checkable claims from what it pulled.
A "Salesforce vs HubSpot" page announces, in the title, that it contains a structured comparison of two named products. That is a direct match to a head-to-head buying question and a partial match to four of the other five buying archetypes. A "Salesforce review" page announces one product and an opinion, which is a match to almost nothing a purchase-intent buyer asks.
The archetype data reinforces it. Head-to-head comparison prompts named a mean of 2.1 brands and took 91.0% of citations from vendor-owned pages. Alternatives-to prompts named 7.8 brands. Category-landscape prompts named 10.9. Every one of those question shapes is served better by a comparison artefact than by a review.
The sites that understood this first
The comparison-site long tail in this corpus – 43 small independent sites taking 53 citations, 14.4% of the total, 3.5 times what the established review aggregators managed – won on exactly this. They publish dated, titled, tabular comparisons of named products with prices, and they publish them fast.
erpresearch.com earned 11 citations across four answers, making it the fifth most-cited domain in the entire study, ahead of Salesforce, Microsoft, Asana and Gartner. ciopages.com earned 10 across four. Neither has meaningful brand recognition. Both publish the page shape the fan-out is looking for.
These sites are not out-writing anyone. They are out-formatting them.
Retrieval is not citation, and format governs both
Getting retrieved gets you into the 44.7 results the model reads per answer. Only 11.09% of unique URLs read ever become one of the 6.1 citations rendered. Format governs both steps, but through different mechanisms.
Retrieval is governed by the title and the declared shape. Does the title name products, declare a comparison, and assert a date?
Citation is governed by extractable, checkable claims inside the page. Citations bind to specific assertions – price points, tier names, feature availability, integration support. A comparison page that is 1,500 words of prose with no table will get retrieved and then discarded, because there is no discrete claim to attach a pill to.
The conversion table makes the point across source types.
| Source type | Retrieval share | Citation share | Conversion |
|---|---|---|---|
| Review aggregator | 1.9% | 4.1% | 28.8% |
| Independent comparison site | 9.3% | 14.4% | 21.3% |
| Editorial media | 3.0% | 4.1% | 18.5% |
| Analyst / market research | 3.2% | 4.4% | 18.4% |
| Vendor-owned | 55.1% | 65.9% | 16.4% |
| Agency / consultancy / partner | 7.7% | 5.7% | 10.2% |
| Community / UGC | 1.2% | 0.0% | 0.0% |
Agency and consultancy content is the cautionary case: 7.7% of retrievals, 5.7% of citations, a 10.2% conversion rate. It gets found and then dropped, because most of it is narrative rather than tabular. Community content is the extreme version at 0.0% – 31 retrievals, zero citations.
A publishing brief that follows from the data
Build comparison and versus pages before anything else. One per meaningful competitor pair, one per major alternative to your own product. Titled with both product names. Dated. Tabular.
Build alternatives-to pages for your own product and for the category leader. 196 retrievals in this corpus, and the archetype that names 7.8 brands per answer – the widest entry point for a challenger.
Make your pricing page a page. 278 retrievals. Public, extractable, tier-named, with what each tier includes. A pricing page that is a lead-capture form contains nothing citable.
Put a table in everything comparative. Prose gets retrieved and discarded. The claim needs to be discrete enough for a source pill to attach to it.
Date anything that makes a comparative claim. 86.3% of dated retrievals carried a current-year date and only 5.0% were more than two years old. Comparative pages compete against sites that restamp on a cycle.
Deprioritise the review format. 17 retrievals out of 2,680. Whatever the value of reviews in your funnel, in this corpus they are not the artefact AI retrieval is looking for.
Do not confuse volume with shape. The residual "other / product page" bucket is 43.3% of retrievals, and much of it is homepages and feature pages doing exactly what they should. The point is not to publish more pages. It is that among purpose-declared pages, one shape dominates and one is invisible.
FAQ
What type of content does ChatGPT retrieve most for software questions?
Comparison and versus pages, at 525 of 2,680 retrieved results - 19.6% of the corpus and the largest identifiable content type. Best-of listicles followed at 371 results and pricing pages at 278. Review pages accounted for 17 results, 0.6%.
Are review pages useful for AI search visibility?
Barely, in this corpus. Only 17 of 2,680 retrieved results were review pages. Review aggregator domains as a whole were retrieved 52 times, though when retrieved they converted to citations at 28.8%, the highest rate of any source type. The problem is discoverability, not trust.
Why do comparison pages get retrieved more than reviews?
Because the title declares a structured comparison of named products, which matches how purchase-intent buyers ask. Head-to-head prompts name 2.1 brands, alternatives-to prompts name 7.8, and category-landscape prompts name 10.9 - all better served by a comparison artefact than by a single-product opinion.
Do I need tables on comparison pages for AI citations?
It helps materially. Retrieval is driven by the title and declared shape, but citation binds to discrete checkable claims such as price points, tier names and feature availability. A prose comparison with no table gets retrieved and then discarded, which is the pattern agency content shows at 10.2% conversion.
Should I build alternatives-to pages for AI search?
Yes, particularly as a challenger. Alternatives-to pages drew 196 retrievals, and the alternatives-to prompt archetype names a mean of 7.8 brands per answer against 2.1 for head-to-head - making it one of the widest entry points into a category shortlist.
Related in this series
- How ChatGPT Shortlists Software Brands, the full report as a PDF
- Forty-Five Reads, Six Links: How ChatGPT Actually Cites Software Brands
- The Sites Beating G2 in ChatGPT: 43 Comparison Sites, 53 Citations, 3.5x the Aggregators
- Recency Is Close to a Filter: 86.3% of What ChatGPT Read Carried a 2026 Date
- Publish Your Prices: Why Gated Pricing Costs You 30 Points of AI Citation Share
About the research. Nathan Mzumara is an organic growth and AI search practitioner. Page types were classified from retrieved-result titles in the ChatGPT response stream. Method and limitations are stated in the pillar report.
Tags
