AI answers name your product, then recommend a competitor

Four studies published in the past week all landed on the same finding, one most B2B software teams are not measuring: appearing in an AI answer, supplying the sources behind that answer, and being the product the answer actually recommends are three separate outcomes.

The clearest case comes from Foundation Labs, which ran six unbranded buyer queries across eight AI engines and collected 348 responses about Brex, the corporate card company. Brex was named in 51.7% of those answers, second only to competitor Ramp at 60.6%. But brex.com supplied only 6.51% of the citations the engines used to build those answers. The aggregators beat every vendor in the category: Nav took 9.78% of citations and NerdWallet 8.63%. Brex ranks #1 organically for "business credit cards for startups"; in the AI Overview sitting above that #1 result, Foundation found the engine drawing on a Reddit discussion and citing Nav.

Mention rate looks healthy while recommendation rate erodes

Foundation's second piece, published three days earlier, shows why that distinction has teeth. Its team put the same revenue-intelligence buying question to ChatGPT and Claude eleven times. Gong appeared in all eleven threads and won the recommendation in three. The other eight went to Apollo, Salesloft with Clari, and HubSpot with Outreach, or settled on nobody. Same question, same vendor set, four different winners.

What decided it was the job the model inferred from an open prompt. Where it read the buyer as coaching new reps, Gong won; where it read prospecting, Apollo won. The mention list stayed stable across every run; only the pick moved. That is an awkward finding for the AI-visibility dashboards now being sold to marketing teams, most of which report share of voice. That number can hold flat for two quarters while the thing you are paid for goes to a competitor.

It also makes single-run testing near worthless. If one prompt produces four winners across eleven runs, checking your brand in ChatGPT once tells you what the engine said once, not what it usually says.

Your own site is a rounding error in the source set

The second convergence is about where the source material comes from, and software fares worse than most categories. Foundation classified 24.6% of the Brex dataset's citations as competitor sources and 21% as earned media — 45.6% of the inputs, against 6.5% from Brex's own pages. Its broader work, it says, puts roughly 90% of AI citations for B2B SaaS outside the cited brand's website.

Research from Trendos, published as sponsored content on Search Engine Journal and worth reading with that commercial interest in mind, puts a number on the software-specific case. Across a claimed 107 million AI answers, the top ten citation sources in IT and solutions services split 51% community and user-generated content, 47% independent editorial and reference, 2% brand-owned. The editorial half is largely review directories — Trendos names G2, Clutch, Gartner, GoodFirms, SourceForge and Slashdot.

A separate analysis by Shero Commerce, reported by Search Engine Journal, found 2.8% of 1,851 sources cited across Google AI Mode, ChatGPT and Perplexity were brand-owned. Shero is a Shopify agency sampling consumer products, so treat that as corroboration of direction, not magnitude.

A vendor is not going to say 'don't buy this, you're too small for this.' But Reddit will.

That is Troi Leemuel Lamboon, a Reddit specialist on Foundation's strategy team, in the Brex breakdown, on why community threads outperform vendor pages on recommendation queries. Vendor pages are structurally incapable of naming a poor fit, and poor-fit information is exactly what a model needs to separate two products with near-identical feature lists.

What a software marketer should change

The practical consequence is a reporting change before it is a budget change. If under 7% of the inputs shaping an answer come from your own domain, a content plan built almost entirely around that domain is pulling the smallest lever available.

• Add two columns to your search reporting: which company the answer recommends, and which domains it cited. Rankings and mention rate cannot show either.
• Run unbranded, discovery-stage prompts. Naming yourself inflates the result — Foundation excluded its two branded queries from the visibility measurement for that reason.
• Repeat each prompt across engines and over time. Treat a single run as an anecdote.
• Audit the recurring sources for queries you lose. Foundation found Zapier in 70% of source records for the revenue-intelligence question, purely for publishing category roundups.
• Fill in the review directories your buyers shortlist from, and chase trade coverage that names the product, before commissioning another landing page.

Two caveats. These are small samples run by companies selling services against the problem they describe, and the engines shift — Foundation says plainly that its findings represent an August 2026 measurement window, not a permanent result. Nobody has yet shown that changing your source footprint changes the recommendation. What the week's research does establish is narrower and still useful: the metric most teams report is not the metric that decides the shortlist.