I Put 63 Buying Questions to ChatGPT and G2 Was Cited Zero Times

The advice is everywhere and it is always phrased with total confidence: if you want ChatGPT to recommend your product, get reviews on G2 and Capterra, because that is where the model reads about software. I have seen it in agency decks, in LinkedIn posts, and in at least three AEO vendor blogs this year. It sounds right. G2 ranks for everything, the pages are structured, and the content is exactly the comparison data a model would want.

Nobody in that pile had actually counted. So on 8 and 9 September I counted.

Download the dataset (CSV, 95 rows) — every corpus, the domains that showed up in it, how many citations each got, and what kind of page it was.

What I ran

I had a list of 40 B2B SaaS companies I was scanning for outbound anyway. For each one, our scanner writes three buying questions about that company's own category and puts them to ChatGPT with search turned on, then keeps every URL the answer was built from. Not what the answer said. The sources under it.

The questions are the ordinary kind:

• "What's the best CRM software for small to medium businesses in 2024?"
• "What are the top alternatives to Salesforce that are more affordable?"
• "What's the best sales tracking app for field sales teams that need automated reporting?"

These are precisely the searches G2 is supposed to own. Somebody comparing CRMs, asking for alternatives, asking what is best for their size of company. If review aggregators are the substrate of AI recommendations, this is where they would show up.

Nineteen of the 40 came back with a usable corpus. The rest either failed the crawl or returned an answer with no citations at all, which is its own problem and not the one I am writing about today. That left 57 questions and 294 citations across 19 categories, mostly CRM and business software, with some marketing and field sales.

The result

techradar.com44(13 of 19 corpora)

close.com5(5 corpora)

analyticsinsight.net4(3 corpora)

blog.hubspot.com3(3 corpora)

prnewswire.com3(2 corpora)

capterra.com1(1 corpus)

g2.com0(never cited)

trustpilot.com0(never cited)

getapp.com0(never cited)

softwareadvice.com0(never cited)

gartner.com0(never cited)

Citations, ChatGPT with search. 57 buying questions, 19 corpora, 8-9 September 2026. Counted from the top five domains in each corpus, which hold 142 of the 294 citations collected.

G2 did not appear once. Neither did Trustpilot, GetApp, Software Advice or Gartner. Capterra managed a single citation, in one CRM corpus, on one question.

The absence is more striking when you look at what kind of pages did get through.

Article103

Listicle29

Comparison2

Review site1

Citations by page type. The 142 classified citations. One further page classified as a review site was Wikipedia, which our classifier got wrong, so it is excluded here.

One citation out of 142 came from a software review site. Ordinary articles, written by publishers, took 103 of them.

It is not a robots.txt problem

The first thing anyone says when they see this is that the review sites must be blocking the crawlers. It is the obvious explanation and it has been the correct one before. When Reddit's presence in ChatGPT collapsed in August, half the internet blamed robots.txt, and that turned out to be wrong too, because the block came after the drop.

So I read the files, on 12 September 2026.

G2's robots.txt gives GPTBot exactly the same rules as everyone else, a list of tracking-parameter and login paths, nothing more. It does not mention OAI-SearchBot, the crawler that actually fetches pages for ChatGPT's search, which means that bot falls under the permissive wildcard rule at the top of the file.

Capterra goes further and names the bots one at a time:

User-agent: GPTBot
Allow: /

User-agent: ChatGPT-User
Allow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

That is a site that has thought about AI crawlers and decided to let all of them in. It got one citation out of 294.

Trustpilot allows OAI-SearchBot everywhere except a few review-submission paths like /evaluate/. Also readable, also absent.

All three are open. The model can reach them and did not use them. Whatever is happening here is happening inside the retrieval and ranking step, not at the door.

What was there instead

techradar.com, over and over. Forty-four citations, present in 13 of the 19 corpora, and in most of those it was the biggest single source in the answer.

The URLs are all the same shape. techradar.com/reviews/sugarcrm-review. techradar.com/reviews/insightly-crm-review. techradar.com/reviews/bitrix24-crm-review. A UK consumer-tech publication, better known to most people for phone and laptop reviews, is currently answering "what's the best CRM" for a good chunk of the B2B software market.

I noticed this first as a footnote in an outbound batch and assumed it was a quirk of one list. It was not. I have now seen the same domain come back at the top of a category corpus in two separate runs, on two different lists, on two different days.

If you want a guess at why, and it is only a guess: a TechRadar review is one product, one URL, several hundred words of prose with a verdict in it. A G2 category page is a sorted, filtered, paginated grid of 200 products where the ranking is generated from review volume. The first is an answer to the question. The second is a tool for a human to find an answer with, and a model that has already decided what to say needs a sentence it can point at.

The corpora are thinner than you would expect

The other thing that surprised me is how small these source pools are.

Across the 19 corpora, three questions produced an average of 15 citations drawn from 13 distinct domains. In total there were 247 domain slots holding 294 citations, so the typical domain in a typical corpus was cited exactly once and never again.

Concentration varied wildly. In the tightest corpus the top five domains held 82% of all citations; in the loosest they held 29%. That range matters more than the average, because it tells you the game is different in different categories. Where five domains hold four fifths of the answer, getting into one of them is the whole job. Where the answer is spread across 18 domains cited once each, there is nothing stable to buy your way into, and you are better off owning the question on your own site.

The rest of the corpus is stranger than the review-site model would predict. Vendor blogs get cited about their own competitors, which is how close.com ended up cited in five different corpora. A Brazilian software directory showed up twice. One answer leaned on a press release from PR Newswire.

Then I checked my own category

A day later I ran the same thing against cuescout.com with buyer-shaped keywords for my own category: AI visibility tools, AEO tools, Profound alternatives, cheap AEO tool for small teams. Six questions, 28 citations.

G2, Capterra and Trustpilot: zero, again.

What was there: techradar.com at the top for the third time, ahrefs.com/blog, a Webflow app-directory listing for a competitor, an independent ranking site called aeo-rankings.com that I had never heard of, and a competitor's own media-kit page. Also a German staff-scheduling company and a cancer research consortium, because the corpus for a young category is genuinely half noise.

That was the run that killed our affiliate program, incidentally. We had a list of 20 publishers who rank on Google for "best AEO tools" and were about to pay them 30% recurring. None of the 20 were in the corpus the model actually reads. Google rankings and model citations turned out to be two different maps of the same category, and we had been navigating with the wrong one.

Where this is thin

I want to be blunt about the limits here, because the headline number is the kind that gets screenshotted.

Our public scan only hands back the top five domains per corpus and locks the rest behind a signup. That means I can account for 142 of the 294 citations. G2 could be down in the tail with one citation somewhere. What I can say is that across 19 corpora it was never in the top five, and the tail is where domains sit that got cited once by one question.

One engine. This is ChatGPT with search, nothing else. When we ran 36 questions across three engines in August, Perplexity cited Reddit 38 times and ChatGPT cited it zero times out of 139, on identical questions. The engines read different internets, and it is entirely possible Perplexity loves G2 and I have not checked.

The sample is also lopsided toward CRM, because the list I was scanning was lopsided toward CRM. A run on developer tools, HR software or fintech might look nothing like this, and CRM in particular is a category where TechRadar has been publishing reviews for a decade.

Two days of collection. Source mixes move. The Reddit collapse in August took ten days to wipe out 83% of one platform's presence in ChatGPT, so a single run is a photograph, not a trend line.

And these questions were machine-written from each company's website. A real buyer types worse and more specifically than our generator does.

What I would actually do with this

Stop buying review campaigns and calling it AI visibility. Get reviews for the reasons that were always true, which is that a buyer in a bake-off will go and read them, and your sales team wants the badge for the deck. Take "so ChatGPT will cite us" out of the business case, because in 19 out of 19 categories here it did not.

Then go and find out who is actually in your category's corpus, because it will not be who you expect. For 13 of these 19 companies the answer was a UK consumer-tech site they have almost certainly never pitched. The work that follows is an email to whoever maintains that review, not another review-site campaign. We wrote up how to get into the pages that do get cited separately.

Check the concentration before you spend anything. If five domains hold 80% of your category's answer, go after those five. If the citations are spread thin across eighteen domains cited once each, no amount of outreach will hold, and the money is better spent on pages you own.

The uncomfortable version of this finding is that in most of these categories, the product that wins the answer is not the best-reviewed product. It is the one TechRadar reviewed in 2023.

You can run the same check on your own category with our free source radar. Put in one real buying question, look at the domains that come back, and see whether the sites you have been investing in are in the list. No signup, nothing stored.