Method note: this article reads three public case studies and two citation-structure studies, all linked below. Where the source is a vendor or an agency selling related services, we say so inline. Every number is quoted from the linked page.

A buyer who used to type "best feature flag tools" into Google now asks a chatbot. The answer names three or four tools, and the page of ten blue links never opens. G2's 2026 survey of 1,076 B2B software buyers put a number on the shift: 51% now start product research with an AI chatbot more often than with Google. A year earlier it was 29%.

If your tool is in that answer, you are on the shortlist. If it is not, the buyer never hears about you, because most of them never reach a search results page at all. This article walks through what the public record says about how tools actually get into those answers, using cases that published numbers.

What AI answers draw on when they name tools

A study of 5.17 million links cited by AI answers across industries (Parse) found that for B2B software, a vendor's own site accounts for only 25-29% of its citations. General web content accounts for 46-52%, and community forums 10-13%. A separate analysis of 57.2 million citations by Foundation and AirOps looked at DevOps categories specifically and found answers leaning on GitHub, Medium, and dev.to.

The pattern is blunt: your polished website is a minority of the evidence. Tools that get named tend to have left traces on other people's surfaces, in comparison lists others wrote, in community threads, in tutorials on platforms the engines already trust.

Case one: Rootly, from 3% to 30%

Rootly worked with AthenaHQ, a generative-engine-optimization vendor, and published the results: a citation rate growing from 3% to 30% across the engines they tracked, a tenfold increase, with mentions on unbranded questions up 2.5 times. The case study is vendor-published, so treat the magnitude with care. What it shows is the shape of the work: the gains came from publishing citable evidence, not from changing the product.

Case two: Gumlet, 20% of inbound revenue

Gumlet, a video and image CDN, traces close to 20% of its inbound revenue to AI discovery. The number comes from a case study by DerivateX, the agency that ran the program. It quotes Gumlet's co-founder and CMO, Divyesh Patel, and combines an onboarding survey (which put the share at 25.96% for the week of October 12, 2026) with GA4 revenue data. Agency case studies flatter their authors by default. This one at least shows its measurement path, and 20% of inbound revenue is the kind of number a CFO argues with, so the agency had to be able to defend it.

Case three: DigitalOcean's tutorial library

The oldest and largest case predates AI answers entirely. DigitalOcean paid developers to write task tutorials (how to set up OAuth 2, how to configure a server) and reached roughly 6 million monthly visits from about 1,400 tutorials, with 69% of its customers reading them. The figures come from cofounder Mitch Wainer's 2016 talk at Heavybit. The mechanism is exactly what current citation studies reward: tutorials answer the task the buyer is trying to complete, the product stays in the footer, and the pages accumulate on a domain the engines have learned to trust.

What the three cases share

  • They publish answers to the questions buyers actually ask, in the buyers' own words.
  • The evidence lives partly outside their own domain: comparisons, community, third-party tutorials.
  • None of it is a config file. Our audit of 11 incident management tools found tools with textbook llms.txt setups sitting at 0% presence. Open doors are entry requirements, not levers.

Questions your buyers are already asking

Illustrative examples, phrased the way buyers phrase tool searches. Swap in your own category:

  • "Best open source feature flag tools for a small team"
  • "What's a good self-hosted alternative to [commercial product]?"
  • "Which monitoring tool is easiest to set up for a Node.js app?"

Run your list against ChatGPT, Perplexity, and Gemini and record whether your tool comes back. If you would rather have it done for you, CiteGap's free scan runs 5 questions on 1 engine. The written diagnosis ($490) covers 30 real buyer questions across all three engines and returns an engineering-grade fix list. For the devtool-specific playbook, see AI visibility for devtools.

Get the written diagnosis ($490)

Frequently asked questions

How do AI engines decide which devtools to recommend?

They compose answers from citable evidence rather than ranking vendors. For B2B software, the Parse study found a vendor's own site contributes 25-29% of its citations, with general web content at 46-52% and community forums at 10-13%. There is no submission step: engines repeat what has been written where they read.

Do the case study numbers hold up?

With caveats we state inline. Rootly's tenfold citation increase is vendor-reported by AthenaHQ, which sells the service. Gumlet's 20% figure comes from the agency that ran the program but shows its measurement path (survey plus GA4). DigitalOcean's figures are first-hand but date to 2016. Treat the magnitudes as directional and the mechanisms as the takeaway.

Can a small tool without a content team do any of this?

Yes, in a narrower form. Pick the ten questions your buyers actually ask, publish honest comparison pages that name competitors, and be visibly useful in one or two communities where your category is discussed. That is the minimum evidence pool. You can run the structural checks yourself before writing anything.

How do I check whether AI recommends my tool?

Run your category's buyer questions across ChatGPT, Perplexity, and Gemini and record which tools come back. Or use CiteGap's free scan (5 questions, 1 engine) as a first read, and the written diagnosis for the full 30-question, three-engine version.

Sources