Study · 19 July 2026

We asked three AI assistants 40 buying questions. Reddit fed 95% of the answers.

Everyone in this category writes the same article about how to “optimise for AI search”, and almost nobody publishes what the assistants are actually reading. So we measured it: 40 real software buying questions, put to ChatGPT, Gemini and Perplexity, with every source recorded. The result is uncomfortable if you have been paying a review directory, and clarifying if you have not.

40

buying questions

292

citations recorded

196

distinct domains

95%

of answers used Reddit

How we ran it

We took 40questions of the kind a person actually types when they are about to buy something: “best project management tools for small teams”, “calendly alternatives”, “google analytics alternatives 2026”, “stripe alternatives for saas”. They spread across twenty-odd categories on purpose, from project management and CRM to hosting, auth, payments, invoicing and error tracking, so that a quirk of one market could not carry the result.

Each question got one live web search, and the identical results were handed to ChatGPT and Gemini. Perplexity got the bare question, because it always runs its own retrieval and will not read results handed to it. Then we recorded two things per question: every domain that fed the answer, and every product the answer named. All 40 questions returned an answer. Nothing was retried to get a nicer number, and nothing was excluded.

One detail that changes how you should read every number below, so it goes here rather than in a footnote: we record up to eight sources per answer. The median answer hit that ceiling, so 292is a floor, not a census. What that makes these percentages is arguably more useful than a full count would be. “Reddit appeared in 95% of answers” here means Reddit was in the top handful of sources 95% of the time, which is the part of the retrieval that actually shapes the sentence you get.

Three more honest limits. This is one run on one day, 19 July 2026, in English, on software categories. Retrieval changes hour to hour and the exact percentages will move if you run it tomorrow. A citation is not an endorsement: it means the page was retrieved and read, not that the assistant agreed with it. And the answers were merged across the three assistants, so this describes what they collectively read rather than which one leaned hardest on what. What the numbers are good for is the shape of the thing, and the shape is not close enough to the margin for a rerun to reverse it.

Finding 1: Reddit is not a channel, it is the substrate

Reddit appeared in the sources of 38 of the 40 answers. Not 38 mentions across all answers, 38 separate questions where at least one Reddit thread was part of what the assistant read before answering. The two questions that came back without it were “best email marketing software for creators” and “best seo tools for small sites”, both categories with an unusually dense layer of affiliate roundups sitting on top of them.

The gap to second place is the part worth sitting with. YouTube, at 35%, is not close. Reddit is not the biggest source in the way a market leader is biggest. It is the default, and everything else is an occasional supplement to it.

The practical reading of this is not “go and spam Reddit”, which will get you removed and is the one strategy guaranteed to produce a thread that argues against you. It is that the questions your buyers ask have already been asked in public, by someone, in a thread that is now load-bearing infrastructure for every assistant. Find those threads. Being genuinely useful in three of them beats a year of posting on a channel nobody retrieves.

Finding 2: the review sites you pay for are barely there

This is the finding we did not expect to be so stark. Across all 292 citations, the five classic software review directories accounted for 0.7%.

G2 was read for two questions. Capterra, TrustRadius, GetApp and SoftwareAdvice were not read for any of the forty. Meanwhile Reddit alone accounted for roughly 13% of every citation in the study, which is about nineteen times the combined share of the directories.

Be careful with what this does and does not say. Review sites still sell: they rank in ordinary search, they carry buyer intent, and a category leader badge still closes deals with procurement teams. What the data says is narrower and still expensive if you get it wrong. If you are paying a directory because you believe it is how AI finds you, that specific belief is not supported here. The budget is doing something, but not that.

There is a plausible mechanism behind it. Directory pages are largely templated, gated behind interaction, and thin on the actual prose a model needs to form a sentence like “X is good for Y because Z”. A Reddit thread where four people argue about which tool broke at scale is dense with exactly that.

One answer, taken apart

Aggregates hide the mechanism, so here is a single question in full. We asked for “calendly alternatives”. These are the six sources the answer was built from, in the order they were retrieved:

• calendly-alternatives.orga domain that exists to be exactly this answer
• reddit.comthe thread where people argue about it
• lunacal.aia scheduling product's own site
• emailtooltester.coman affiliate roundup site
• quora.coma second community source
• spectroomz.comanother vendor's own site

And the eight products it named: Cal.com, Doodle, HubSpot Meetings, OnceHub, YouCanBookMe, Acuity Scheduling, Koalendar, TidyCal.

Look at what is not there. No G2. No Capterra. No software-review authority of any kind. Instead: one purpose-built comparison domain, two community threads, one affiliate site, and two scheduling products that got into the source list on the strength of their own website. Neither Lunacal nor Spectroomz is a household name. They are in the answer because their pages were retrievable and relevant when the question was asked.

Finding 3: you can be your own source, and most winners are

The Calendly example is not a fluke. In 29 of the 40 answers, at least one of the products the assistant named also had its own domain sitting in the source list. Just under three quarters.

This is the finding that should change what you do on Monday, because it contradicts the most common thing people believe about AI visibility. The usual framing is that you are at the mercy of what others write about you, that AI answers are a popularity contest decided elsewhere and all you can do is wait. The data says otherwise. Most of the time, at least one named product is in the answer partly because of a page it controls, wrote, and published itself.

That page has to earn retrieval, which means it has to be the kind of page that answers the question rather than sells at the person asking it. A landing page full of “transform your workflow” will not do it. A page that plainly says what the product is, who it is for, how it compares to the obvious alternative, and what it costs, will. The bar is lower than the industry implies and higher than a hero section.

Finding 4: YouTube is a text surface now

YouTube fed 35% of the answers, which surprised us enough that we checked the rows by hand. It holds up. Assistants are reading transcripts, descriptions and comment threads, and a fifteen-minute “I tried six CRMs” video is, to a retrieval system, a long text document with a strong opinion and a clear list in it.

For a small team this is the most under-priced surface in the study. Competition for a written “best X” roundup is brutal. Competition for an honest, specific video on the same topic is meaningfully lower, and the thing that gets retrieved is the words, not the production quality.

Finding 5: a vendor blog beat every review site

Zapier’s blog fed 30% of the answers, third overall. It is not a neutral source. It is a software company’s content marketing, and it is being read by assistants roughly six times more often than G2.

Zapier earned that with volume and genuine usefulness over many years, which is not a plan you can start this quarter. But the shape of what worked is copyable at any size: exhaustive, specific, comparison-shaped pages that answer the question completely instead of teasing a demo. Every third-party “best X” page in the top of this table has the same structure. The models are not rewarding brand, they are rewarding pages that resolve the question.

The two categories where the community vanished

Reddit was in 38 of 40 answers. The two exceptions are worth more attention than their size suggests, because they were not random: “best email marketing software for creators” and “best seo tools for small sites”. Both came back with zero community sources. Not one Reddit thread, not one Quora answer, in either.

Those are two of the most heavily monetised affiliate categories on the internet. Email tools and SEO tools pay recurring commissions, so a dense layer of professionally optimised roundup sites has been built on top of every query in them, and that layer outranks the discussion. The assistant is not choosing to ignore real users. It is retrieving the top results, and in these categories the top results are owned.

This has a direct consequence if you sell in a category like that. Community strategy will underperform for you relative to what everyone else is advised to do, because the community is not what gets retrieved in your market. Your equivalent move is to be in the roundups, which means the unsexy work of actually contacting the sites that publish them, being listed accurately, and being worth listing. It is slower and less pleasant than posting on Reddit, and in these two categories it is the thing that works.

Across the rest of the study the community share of sources ranged up to about a third per question. So the honest instruction is not “go to Reddit” universally. It is: measure your own category first, because the answer differs enough between categories to send you in the wrong direction for a quarter.

Who actually got named

266 distinct products came up, and the concentration at the top is much weaker than you would guess. Framer was the single most-named product in the entire study, and it appeared in three questions out of 40. Everything else that repeated at all appeared in exactly two: Asana, Trello, ClickUp, Jira, Airtable, Brevo, Plausible, Fathom, Matomo, PostHog, Typeform, Jotform, Netlify, Render, DigitalOcean, Railway, Paddle, FastSpring, Semrush, Ahrefs, n8n, Make, and about twenty others.

Two things stand out. The first is how many of these are not the category giant. Fathom and Plausible are small analytics companies that appear as often as anything else in their category. Dodo Payments and PayPro Global sit in the same list as Paddle. Being named is clearly not a function of company size, which is the most encouraging sentence in this article if you are the small one.

The second is that there is no dominant name anywhere. Nobody was named in more than three of forty questions. If you have been treating AI visibility as a race with an unreachable leader, the data says there is no leader to reach. There is a wide, shallow field where almost everyone is named occasionally and almost nobody is named consistently, and consistency is therefore available to whoever bothers to work at it.

Finding 6: the head is small and the tail is enormous

292 citations came from 196 distinct domains, and 173 of those domains appeared exactly once. That is 88% of the domains contributing a single citation each. The top four together held only 24.7% of all citations.

Both halves matter. The head means there are four or five places you cannot afford to be absent from. The tail means the remaining three quarters of the answer is assembled from a churn of niche blogs, forum posts, developer sites and one-off comparisons that changes per question. You cannot buy your way into that tail. You get into it by existing, in text, in enough specific places that a retrieval pass keeps bumping into you.

This is also why a single mention feels like it does nothing. It genuinely does almost nothing. It is one ticket in a draw that gets re-run on every question.

Finding 7: being named once is not a position

The answers named 266 distinct products, an average of 7.75 per question. Of those, 223 were named in exactly one question and only 43 were named in more than one. No product was named in more than three.

So the shortlist is not a stable ranking that you climb. It is re-drawn per question, out of whatever the retrieval pass happened to surface. Two consequences follow, and they are the two most useful sentences in this article.

First, checking one question tells you almost nothing about your visibility. If you ask ChatGPT once, see your name, and conclude you are fine, you have measured a coin flip. Ask the ten questions your buyers actually ask, and count.

Second, the winnable goal is not “be number one”. It is to be one of the 7.75names that comes up, across as many of your category’s questions as possible. That is a far more achievable target than the language around AI visibility usually implies, and it is achievable by being thoroughly present rather than by being the biggest.

What this means for the SEO you are already doing

The loudest claim in this industry right now is that AI search makes SEO obsolete and you need a separate discipline for it. This data does not support that, and the distinction it does support is more useful.

Every source in the study is an ordinary web page that an ordinary crawler can reach. There is no secret AI index. When an assistant answers “best CRM for solo founders”, it runs something very like a search, takes something very like the top results, and writes a paragraph from them. Which means the machinery you already own, crawlable pages, clear titles, internal links, content that ranks, is the same machinery that decides whether you are in an AI answer. The three days of SEO work we describe in our case study moved our traffic 64%, and it is the same work that makes a page retrievable.

What transfers directly

Crawlability, page speed to the extent it affects crawling, clear title and heading structure, internal linking, and being genuinely useful enough to rank. If a page cannot be fetched and parsed, it cannot be retrieved, and nothing else you do matters. This is the unglamorous 80%.

What is genuinely different

Three things, and they are worth internalising because they change what you optimise for.

There is no position two. In search you can be fifth and still get clicks. In an answer you are named or you are absent, and the average answer here named 7.75 products out of a category that might contain fifty. The distribution is far more brutal than a results page.

Someone else’s page can win it for you. Ranking is about your pages. Being named is about any page, anywhere, that the assistant retrieves. A Reddit thread you did not write and cannot edit is doing more for some products in this study than their entire content programme. You influence that, but you do not own it.

The unit of measurement is the question, not the keyword. Keyword rank is stable enough to check monthly. Whether you get named is re-decided on every question, by whatever retrieval surfaces that time, which is exactly why 223 of the 266 products here appeared in only one question. You have to sample across questions or you are not measuring anything.

So the honest summary is not “SEO is dead”. It is that GEO is mostly SEO plus a presence problem you cannot solve on your own domain, measured in a way that ordinary rank tracking cannot see.

Run this yourself

We would rather you checked us than believed us, and the method is deliberately simple enough to repeat by hand. Write down the ten questions your buyers actually ask, phrased the way they would type them, never including your own product name, because a question containing your name guarantees an answer containing your name and teaches you nothing.

Put each one to ChatGPT, Gemini and Perplexity with web search on. For each answer, write down two things: every source it cites, and every product it names. Then count. Ten questions across three assistants is an afternoon, and at the end of it you will know your real coverage, the domains deciding your category, and which competitors keep appearing where you do not.

That afternoon is worth more than any article about GEO, including this one. It is also exactly what we automated, because we got tired of doing it in a spreadsheet.

What we would actually do with this

In the order we would do it, given a normal amount of time and no budget:

Where this study is weak

We would rather say this ourselves than have it said back to us. 40 questions is enough to see a shape and not enough to put a confidence interval on any single percentage. One day of retrieval is a snapshot of a system that changes continuously. English-language, software-category questions are not the whole world, and we would expect local-services or physical-product categories to look quite different.

The per-model split is also missing here. We recorded the merged panel answer, so this study says what the three assistants collectively read, not which of the three leaned hardest on Reddit. That is the obvious next study and we will run it.

What we are confident about, because the margin is not close: community discussion is the dominant retrieval substrate for software recommendations, classic review directories are close to absent from it, and a single mention of your product is not a position. If any of those three turns out to be wrong, it will not be by a little.

Related reading

The strategy this sits inside is in our GEO guide, the tactical version is in how to get mentioned in AI search, and our own numbers, including the unflattering ones, are in the 61-day case study.