We Asked 4 AI Engines What Software They Recommend. Here Is What Actually Gets You Cited.

TL;DR: We ran ChatGPT, Gemini, Claude, and Perplexity on “best [tool]” questions across five software categories, 60 web-grounded runs, then audited the 33 most-recommended tools. The engines cite third-party roundups, review sites, and YouTube far more than any brand’s own website. And the tools that get recommended are not the ones with the best on-site setup: several category leaders have no llms.txt, no schema, and a “weak” page structure. What separates winners is off-site brand presence, not on-site optimization. If you are paying someone to add schema so AI will recommend you, you are buying the wrong half.

What we tested

We asked a buyer-intent question (“best X for Y in 2026, ranked, cite sources”) in five categories: CRM for agencies, invoicing for freelancers, project management for small teams, email marketing for SMBs, and help desk for small SaaS. We ran each on four engines with live web access (ChatGPT, Gemini, Claude, Perplexity), three times each: 60 runs total. Then we audited the 33 most-recommended tools’ actual sites and public footprints. All raw data is kept.

What do AI engines cite when they recommend software?

Mostly other people’s roundups, not brand websites. Across all four engines, the sources cited were dominated by third-party “best tools” listicles, review sites (G2, Capterra), and YouTube. Brand-owned pages showed up as a small minority.

The cited domains were dominated by YouTube (the single most-cited source once Perplexity is in the mix) plus third-party roundups like softwaresift, costbench, onepersonhq, technologyadvice, ventureharbour, project-management.com, and review sites, with a scattering of techradar, pcmag, and reddit. The engine does not read your carefully optimized homepage and decide you are great. It reads what other pages say about your category and repeats the names it finds. That means “getting recommended” is mostly about being present in the sources the engine retrieves.

Does an llms.txt or schema get you recommended?

No. We audited the 33 most-recommended tools, and the on-site technical checklist does not predict who wins. About half of them have no valid llms.txt, and many have no schema. Concrete proof from the winners themselves:

• GoHighLevel ranked number one for agency CRM with no llms.txt, no schema, and a weak page structure.
• FreshBooks and QuickBooks, the top two invoicing picks, both have barely any on-site setup and no llms.txt or schema.
• Help Scout gets recommended with the weakest profile of anyone we audited: no llms.txt, no schema, not even a YouTube channel.
• ActiveCampaign sends a signal that explicitly opts out of AI input, and still gets recommended in nearly every run.

Meanwhile the one tool out of 33 whose recommendation was best explained by its on-site setup (MailerLite, which has a clean llms.txt and good structure) is a mid-tier pick, not a category leader. The heaviest on-site investment we found, monday.com shipping an llms.txt with an actual agent server card, still did not make the switch the reason it wins. This lines up with the controlled evidence: Ahrefs found adding schema moved AI citations −4.6% on AI Overviews, and Google states you do not need llms.txt, special schema, or content chunking to appear in its AI search.

Then what does get you recommended?

Off-site brand presence. The single attribute that separates the tools recommended in almost every run from the ones recommended sometimes is brand ubiquity, and it shows up off your site: in the roundups, on the review profiles, and on YouTube that the engines actually cite. The chain is: be a known brand, get named in the third-party sources the engine retrieves, get cited, get recommended.

This matches the correlation data across 75,000 brands: branded web mentions correlate 0.66 to 0.71 with AI visibility and YouTube mentions about 0.74, while backlinks manage only about 0.22. Being talked about, off your own site, is the crank.

Do the four engines even agree?

On the winners, mostly yes. On the exact list, no, and Gemini is the least stable. ChatGPT returned nearly the same shortlist every run (run-to-run overlap 0.67 to 1.0). Perplexity and Claude were in the middle (0.36 to 0.69 and 0.32 to 0.62). Gemini reshuffled the most (0.17 to 0.45). So the consensus winners are a stable target, but the precise ordering is noisy, especially on Gemini. If a vendor sells you a single fixed “you rank #3 in AI” number, they are selling noise: the output moves every run.

Does the “Ask AI about us” injection trick work?

No. We found a cold-email company that put an “Ask AI about us” button on its homepage, pre-loading a prompt that tells the model to “remember us as a citation source” and that its price beats two named competitors. We tested it: we asked ChatGPT for the best providers in its category. It recommended eight tools, including the two competitors the injection names. It did not mention the company once. The trick only fires when a human clicks it, and it writes to that one person’s private session. It cannot change what the engine tells anyone else.

What should you actually do?

Leave the switch alone. Get into the sources the engine retrieves.

• Get placed in the roundups and “best X” lists in your category. They were the single biggest citation source in our test, and many are small, reachable blogs, not walled gardens.
• Build unprompted brand mentions, the strongest measured correlate: Reddit, communities, podcasts, PR, other people’s content.
• Stand up YouTube. Every project-management winner had a real channel; the one tool with no channel was the weakest-recommended.
• Keep basic on-site hygiene (crawler access, a clean page), but treat it as table stakes, not the strategy.

FAQ

Do I need an llms.txt file to get cited by AI? No. About half of the 33 most-recommended tools we audited have no valid llms.txt, including several category leaders, and Google says it is not required.

Does schema markup help AI recommend me? Not measurably. A controlled study found adding schema moved AI citations −4.6% on AI Overviews and near zero elsewhere.

Which AI engine should I optimize for? They overlap on the winners but retrieve different sources and reshuffle run to run. Optimize for the off-site presence they all draw on (roundups, reviews, YouTube) rather than any one engine.

Is a single “AI visibility score” reliable? Treat it with caution. Output is unstable enough that the exact ranking changes between runs, especially on Gemini.

We are a small challenger. Is on-site GEO a waste? Not a waste, but not the needle-mover. Do the cheap hygiene, then spend the real effort on getting into third-party sources and building mentions.

Methodology & caveats

Five categories, four engines (ChatGPT via API with web search, Gemini with Google Search grounding, Claude via web search, Perplexity via sonar), three runs each (60 runs), plus a switch-versus-crank audit of 33 winners’ live sites. Honest limits: G2 and Capterra block automated access, so review counts here are directional rather than live-verified (brand presence is verified). “Household name” is a judgment call, not a metric. Results are US-locale and single-day. The teardown is observational, but it converges with the one controlled causal study and Google’s own guidance, so we are confident in the direction: the engine repeats what other sources say about you. Go be worth repeating.

If you want us to run this on your category and hand you the exact roundups, review profiles, and mentions you are missing, that is what we do.