Loading…
Sources field guide
We asked AI 57 software buying questions. 93% of the pages it cited were “best of” lists.
Original data: Claude answered 57 buyer questions with web search. 93% of cited pages were listicles, 52% sat on vendors’ own sites, and none were forums.
By AnswerMentionsReviewed Oct 17, 20268 minute read
Written for: Founders and marketers at software companies deciding where effort will actually move AI recommendations.
Short answer
When Claude recommended software with web search, 93% of the pages it cited were “best X” or “X alternatives” lists, and 88% put “2026” in the title. Software vendors’ own sites supplied 52% of citations, independent review sites 33% and forums none. A vendor whose own list was cited got named in the answer 35% of the time.
Our position
For software, the “best X for Y” list is now the unit of AI visibility. Not your homepage and not your schema: the lists AI reads seconds before it answers. Write one worth citing, get onto the independent ones, or let your competitors’ lists decide your category’s shortlist.
What you should leave with
- 93% of cited pages were listicles or comparisons, and 88% carried “2026” in the title or URL.
- Vendors’ own sites supplied 52% of citations, review marketplaces such as G2 3%, forums and video 0%.
- A vendor’s cited list named its author 35% of the time; the other 65% it sent the buyer to rivals.
- Rerunning a question kept most brands (overlap 0.73) but changed the first pick about one time in three.
How we ran the study
We asked Claude Haiku 4.5, through Anthropic’s API with its web search tool, 60 buyer questions across 12 small-business software categories from a US location on October 17, 2026. Each question ran once with search optional and once with search required, and 36 ran a second time. 57 of 60 completed in each mode.
The questions were written the way a buyer asks before knowing any vendor: the best CRM for a five-person sales team, the cheapest email tool that is still good, alternatives to Gusto. There were five per category: CRM, project management, email marketing, accounting, website builders, payroll, help desk, password managers, SEO tools, ecommerce, scheduling and applicant tracking.
Claude decides when to search from the prompt and returns its sources as citations. We counted every site Claude cited, once per answer, and classified all 227 cited domains by hand as vendors, independent review sites, media, review marketplaces, services firms or forums. Brands were matched against a dictionary built from the answers themselves. The full dataset, with every answer, is downloadable as a CSV.
Method boundary: One model, one day, one location, through the API rather than the Claude app. ChatGPT, Perplexity and Google’s AI features use different search back ends, so their source mix will differ. Read this as a careful snapshot, not a law.
Finding 1: AI cites lists, not product pages
Of 259 cited pages, 242 (93%) were listicles or comparisons: “best X for Y”, “X alternatives”, “A vs B”. Product pages, pricing pages and documentation barely appeared. And 227 of the 259 (88%) carried “2026” in the title or URL, so a dated, current list beat an evergreen page almost every time.
This follows from how buyers ask. “What’s the best CRM for a real estate agent?” is answered fastest by a page that already ranks options for that buyer, so that is what search returns and what the model quotes. Each answer ran about one search, read about nine results and cited 4.5 sites.
Lists shaped the answers directly. For a five-person sales team, Claude named Pipeline CRM second, and Pipeline CRM’s own “best sales CRM software” list was the most-cited page in that answer, three of seven citations.
Finding 2: Vendors write half of what AI cites
Software vendors’ own sites supplied 135 of 259 citations (52%), and 91% of answers cited at least one vendor page. Independent review and affiliate sites supplied 33%, services firms 5%, media 5% and review marketplaces such as G2 3%. Forums, Reddit and video supplied none in this sample.
The mix depends on the category. In help desk, scheduling, ecommerce and project management, vendors supplied most citations. In payroll, password managers and SEO tools, independent review sites did.
This is narrower than wider studies. Muck Rack’s May 2026 study of more than 25 million links cited by ChatGPT, Claude and Gemini found earned media made up 84% of citations across many kinds of questions. Software shortlists are the corner where vendors write the most lists.
| Source type | Share of citations | Answers citing at least one | Examples |
|---|---|---|---|
| Vendor blogs and guides | 52% | 91% | pipelinecrm.com, brevo.com, ones.com |
| Independent review and affiliate sites | 33% | 84% | fitsmallbusiness.com, emailvendorselection.com |
| Mainstream and trade media | 5% | 21% | forbes.com, techradar.com, nerdwallet.com |
| Services firms | 5% | 18% | Bookkeeping firms, SEO and ecommerce agencies |
| Review marketplaces | 3% | 11% | g2.com, trustpilot.com, spotsaas.com |
| Forums and video | 0% | 0% | None |
Free check
See which brands AI recommends in your market
Enter your website or brand. The free audit asks AI the questions your buyers ask and shows who gets named instead of you.
Finding 3: Writing the list doesn’t put you in the answer
When a vendor’s own list was cited, Claude named that vendor in the same answer 47 times out of 135 (35%) and named it first only 8 times. The other 65% of the time, the vendor’s page supplied the shortlist and the buyer heard about its rivals.
The wins were real, and often for small brands. For “alternatives to Calendly”, Claude cited lists by Lunacal, Agiled and YouCanBookMe and named all three, with Lunacal second. For payroll across several states, Netchex’s and Rippling’s lists were cited and both brands made the answer.
The misses were just as clear. Constant Contact’s list of Mailchimp alternatives was cited, and the answer recommended Brevo, MailerLite and ActiveCampaign instead. ONES’s guide to construction project management was cited, and the answer named Procore, Buildertrend and Fieldwire.
ChatGPT shows the same pattern: Ahrefs found a publisher’s own list appeared in 34% of ChatGPT software answers that recommended that publisher. Our reading: a list gets cited for its category facts, and the model builds the shortlist from every list it read.
Finding 4: There is no single site to get listed on
The 259 citations came from 193 different domains, and 155 of those were cited exactly once. The ten most-cited domains supplied only 16% of all citations. The sites that matter are specific to your category and buyer, so they have to be found question by question.
That is the case against buying a generic list of “top directories”. Map the sources behind your own buyer questions instead; our Missing Source Map does exactly that for your category.
| Most-cited domain | Answers citing it (of 57) | Type |
|---|---|---|
| fitsmallbusiness.com | 7 | Independent review site |
| emailvendorselection.com | 6 | Independent review site |
| technologyadvice.com | 5 | Independent review site |
| ventureharbour.com | 4 | Independent review site |
| websitebuilderexpert.com | 4 | Independent review site |
| allaboutcookies.org | 4 | Independent review site |
| forbes.com | 3 | Media |
| techradar.com | 3 | Media |
| ones.com | 3 | Vendor |
| joinhomebase.com | 3 | Vendor |
Finding 5: Shortlists held steadier than sources
Asking the same question again kept most of the shortlist: brand overlap between runs averaged 0.73. But the first-named brand changed in 12 of 34 repeats, and only 8 shortlists came back identical. Cited sources moved more, with an overlap of 0.52. With search optional, Claude still searched 84% of the time.
The nine questions answered from memory were broad ones with obvious leaders: the best ecommerce platform for a small business, the best SEO tool, a free help desk. Narrow questions (a hair salon, a real estate agent, contractors in several states) always triggered a search, and that is where smaller brands got their chance.
SparkToro and Gumshoe measured much more variation in ChatGPT and Google’s AI across 2,961 runs, with a different design. The shared lesson stands: one answer is one draw.
What to do with this if you sell software
Treat category lists as the place AI forms its shortlist. Publish an honest “best X for [your buyer]” page where you genuinely fit, keep its facts and date current, earn a place on the independent lists AI cites for your buyer’s questions, and publish fair comparisons against the category leader.
Google’s guidance on writing high-quality reviews asks for first-hand evidence, comparisons and what sets each option apart. The same standard makes a list worth citing for an AI.
- Pick a buyer, not a category. Narrow questions always triggered a search; broad ones often didn’t. “Best CRM for real estate agents” is winnable; “best CRM” mostly isn’t.
- Write the list you’d trust. Include rivals where they are better. In our data a list helped its author only when the answer could justify the pick.
- Date it truthfully. 88% of cited pages showed 2026. Update the facts first, then the date.
- Get onto independent lists. Review and affiliate sites were cited in 84% of answers, and Ahrefs found branded web mentions track AI mentions more closely than backlinks.
- Map your own sources. The sites behind your buyers’ questions differ by category; the free AI visibility audit shows them for yours.
Test your own brand
See the gap before planning the fix.
Ask ChatGPT three real buyer questions about your market and see who it recommends, free, in about 90 seconds.
Questions that change the decision
Frequently asked questions
Does ChatGPT cite the same kinds of pages?
Lists matter there too, but less. Ahrefs found “best X” lists were 43.8% of ChatGPT’s sources for “best X” prompts, and Seer measured list citations falling in early 2026. Our free AI visibility audit records what ChatGPT cites for your own buyer questions.
Should I publish a “best X” list that includes my own product?
Yes, if you can be honest in it. Vendors whose lists were cited were named in the answer 35% of the time. Rank yourself where you truly fit, show your criteria and prices, and include rivals that beat you for some buyers.
Do review sites like G2 matter for AI recommendations?
In this sample, review marketplaces were 3% of Claude’s citations, far behind vendor and independent lists. They may weigh more on other engines and as trust signals, but they were not where Claude built its shortlists.
Can I use the data?
Yes. The CSV has all 148 answers, the brands each one named, every cited URL and our classification of each domain. It is free to reuse under CC BY 4.0: credit AnswerMentions and link to this page.
Primary sources and research
Platform documentation supports factual statements. Where we describe an audit method or prioritization rule, that is AnswerMentions' operating judgment and is labeled as such.
- AnswerMentions: study dataset (CSV, 148 answers)Every question, answer, brand named, cited URL and domain classification behind this page.
- Anthropic: web search tool documentationThe API tool used for every answer: Claude decides when to search from the prompt and returns cited sources.
- Muck Rack: Generative Pulse, May 2026Across 25M+ links cited by ChatGPT, Claude and Gemini, earned media was 84% of citations.
- SparkToro: AIs are highly inconsistent when recommending brands600 volunteers ran 12 prompts 2,961 times; repeated brand lists were rare.
- Google Search Central: write high-quality reviewsGoogle asks reviews and comparisons to show first-hand evidence and what distinguishes each option.
Continue the diagnosis
Related field guides
Sources