Which sources does AI cite for each kind of buyer question? Six weeks of data

How we measured it

The questions are the 40 fixed English buyer questions of our AI Source Index, each asked in one of four shapes: 16 "best X" questions ("Best running shoes for flat feet"), 8 "alternatives to X" ("Alternatives to Notion for team documentation"), 8 head-to-head comparisons ("Slack vs Microsoft Teams for a remote company of 50") and 8 "how do I choose" questions ("How do I choose a standing desk for a home office?"). Every week they go to Perplexity, ChatGPT, Gemini, Claude, Google AI Overviews and Google AI Mode, once each. We used the six runs of 3, 7, 14, 21 and 28 September and 5 October 2026.

• Unit: one engine's answer to one question in one run. A source (domain) counts once per answer, however many of its pages the answer cited.
• Answers per shape over the six runs: 565 "best", 282 "alternatives", 282 comparisons and 281 "how do I choose". Failed checks are left out rather than counted as answers; 29 of Claude's 40 checks failed on 14 September.
• Each rate is a count out of a count, with a 95% Wilson interval where we give one.

YouTube is the source for head-to-head questions

Share of answers citing each platform, by the shape of the question:

• Head-to-head "X vs Y": YouTube 48.6% (137 of 282; interval 42.8–54.4%), Reddit 25.9% (73 of 282).
• "How do I choose": YouTube 25.3% (71 of 281), Reddit 15.7% (44 of 281).
• "Alternatives to X": YouTube 25.2% and Reddit 25.2% (71 of 282 each).
• "Best X": Reddit 20.9% (118 of 565), YouTube 20.2% (114 of 565).

This is the steadiest pattern in the panel. YouTube was the most cited source on comparison questions in all six runs, and in every run it was cited for all 8 of the 8 comparison questions by at least one engine. It was also the most cited source on "how do I choose" questions in all six runs. Two AI Overviews answers cited YouTube's mobile address, m.youtube.com, which the index keeps as a separate domain; counting them, the comparison figure is 49.3%.

Reddit and YouTube split "best" and "alternatives", and we got one claim wrong

On "best" and "alternatives" questions the two platforms are within a point of each other over the six runs. Which one leads changed during the period: YouTube was the most cited source for both shapes on 3 and 7 September. From 14 September Reddit led "best" questions in all four runs and "alternatives" in three of them, tying YouTube 12 to 12 on 28 September.

On 23 September we said in public, from the 21 September run alone, that Reddit was strong on "best" questions and weak on "alternatives" (15 of 16 "best" questions against 6 of 8 "alternatives"). Six runs do not support that: Reddit was cited for all 8 "alternatives" questions on 3 and 7 September. What does hold is that Reddit's weakest shape is "how do I choose", in every run (4 to 6 of the 8 questions), tied with another shape in two of them. With one answer per engine per question per week, a two-question swing on an eight-question shape can look like a pattern for a week.

The page type follows the question

Share of cited pages whose address says vs, versus, alternative(s), compare or comparison:

• Head-to-head questions: 1,331 of 1,920 cited pages, 69.3%.
• "Alternatives to X": 1,091 of 2,076, 52.6%.
• "Best X": 116 of 3,964, 2.9%.
• "How do I choose": 37 of 1,802, 2.1%.

A comparison or alternatives page competes for comparison and alternatives questions and is close to invisible for "best" questions, which draw on forums, videos and publishers instead: Reddit, YouTube, TechRadar, NerdWallet, CNBC and Forbes lead that list. If a page is meant to win one kind of question, its shape has to match that question. We read the address, not the rendered page, so a comparison page with a plain URL is counted as not comparison-shaped.

Some sources specialise in one kind of question

Consumer Reports was cited 30 times over the six runs, for 4 different questions, and every one of those citations was on a "how do I choose" question: choosing a robot vacuum for a home with pets, a standing desk, a carry-on suitcase and a mattress for back pain. It was cited in all six runs and never once for any other shape.

The four shapes do not hold the same mix of categories, so a pattern like this can be a category effect. The panel has a partial check: the same kind of consumer product is asked in other shapes ("alternatives to Dyson cordless vacuums", the best noise-cancelling headphones, the best running shoes, the best budget monitor, the best fitness tracker), and Consumer Reports was cited for none of them in six weeks.

Among the 43 sources cited at least 12 times for at least three different questions, others lean the same way without that check: Bankrate's 22 citations were all on "best" questions, 21 of Engadget's 23 were on "alternatives", and 23 of Healthline's 31 were on comparisons.

"How do I choose" answers cite nothing most often

24.6% of the answers to "how do I choose" questions cited no source at all (69 of 281), against 10.6% for "alternatives", 9.6% for "best" and 6.4% for comparisons. Two engines account for most of it: Google AI Mode answered those questions without a source 36 times out of 48 and Claude 25 out of 41, while ChatGPT did so once in 48 and Perplexity never. For a buying guide, being the cited source matters most on the engines that cite at all.

What to do with this

• Sort the questions you want to win by shape before deciding what to publish. The same topic asked as "X vs Y" and as "best X" draws on different sources.
• For head-to-head questions, YouTube is cited in about half of the answers, so a video that answers the comparison belongs in the plan, and most of the other pages cited are comparison-shaped.
• For "best X" questions, comparison-shaped pages are rarely cited; the sources are forums, videos and publishers, so earning a place in them is the work.
• For "how do I choose" questions, look at which engines your buyers use: two of the six often answer without any source.
• Judge any of this over several weeks. One run moved Reddit's apparent strengths by a whole shape.

For how long a citation lasts once it appears, see how long an AI citation lasts. For which kinds of sites the engines cite overall, see what AI engines actually cite.

Frequently asked questions