A buyer at a small agency asks ChatGPT which CRM to get, and the answer names a few tools in order. It looks like a ranking, but no stored list of products with scores beside them exists for the model to read out. Keep that one idea, because everything vendors misjudge about AI recommendations starts with picturing such a list.
Each answer is built fresh in four stages: the model sizes up the category, gathers evidence, holds something close to a vote, and orders the result. Each stage answers to different work, so I'll take them in turn for B2B SaaS, with what practitioners have found and where products get in or drop out.
Stage 1: The frame, where new products quietly lose
This stage runs before any search does. The model checks what it already carries about your category: the products it has heard of and the job it thinks each one does. That's where incumbents get their head start, and people who test AI product recommendations systematically keep hitting the same mechanism:
What that testing shows most clearly is that the model isn't snubbing you. It has nowhere to file your product, so it grabs the closest competitor it understands. A newer product loses without being judged at all: it never makes the candidate list, and the slot goes to whatever turns up most in the training text.
I find that encouraging, because it makes the fix a writing job. You need text the model will run into that says plainly what you are and who you serve, and that sets you beside the tools buyers compare you with. Practitioners agree the single best-paying move is a well-built comparison page, which hands the model a finished frame that already includes you.
Stage 2: Retrieval, where the evidence gets gathered
When a product question triggers live search, the engine splits it into subqueries aimed at the places buyers evaluate: community threads and category roundups, plus review platforms and head-to-head comparison pages.
Your own site gets a slice, and one B2B SaaS test measured it. Across its 360 queries, vendor sites took about 36% of the citations, with documentation beating its share. Most of what the engine reads about you, though, was written by somebody else.
Rankings get decided during matching. Take the query "best CRM for a 20-person agency." It carries a spec (team size and use case, plus budget and stack), and the engine hunts for claims that fit. So where is your fit written down? Maybe in a review that mentions agencies, or a pricing tier sized for a small team. Maybe in docs covering the integration that buyer asked about, or a case study with a team that size.
I call this the unwritten-truth problem. If nothing in the corpus says you meet a constraint, the engine assumes you don't, even when you do. That's why I treat publishing your specifics as a ranking tactic.
How often does ChatGPT mention your brand?
Most founders have no idea. The answer might surprise you.
Stage 3: The vote, where consensus gets counted
After retrieval the engine has a pile of evidence and needs to work out what it agrees on. I picture an election where each source gets a ballot. Ahrefs has the biggest dataset on who wins: its 75,000-site analysis found brand mentions correlating with AI visibility about three times more strongly than backlinks.
Practitioners watching product answers see it too: mentions on other people's sites move recommendations further than schema or polishing your own pages.
Who's talking weighs as much as how many are talking. When an engine cites a community thread, it's quoting a customer explaining a tool in their own words, and astroturfed mentions read wrong next to that while real customer language keeps compounding.
Age matters as well. Fresh reviews naming a specific use case beat a big pile of old, quiet ones, so if your evidence dried up in 2024, you're slowly losing these votes to newer, louder voices, whatever your install base.
Stage 4: The ordering, where position gets decided
Last, the engine writes the answer, and where you sit in it is the final ranking that counts. Whatever gets named first sets the buyer's frame for the rest. Our click data and other people's point the same way: the spots cited at the top take a far bigger share of the clicks an answer sends, and the lower ones barely register.
Position gets settled earlier on. A product the model framed well in stage one moves up, and so does one with plenty of agreeing evidence in the vote. It also helps when what you've written lines up with what the query asked.
Descriptions follow the same logic, since the words an engine uses for you, caveats included, come out of that corpus too. Get described as "for enterprises" often enough and you'll look odd in small-team answers, whatever your homepage says.
A shortlist autopsy, stage by stage
One illustrative answer shows the four stages handing out fates. The query is "best customer support tool for a 30-person SaaS," and four products make the answer.
Product A is the category giant, so it was a lock. Its frame alone got it through stage one, and its evidence runs so deep nothing recent could knock it off the top. Product B, a mid-size tool, won its spot in the vote with a two-year stream of reviews that literally include the phrase "small SaaS team," which matched the buyer's constraint during retrieval too.
Product C is newer and scraped in on retrieval by itself. Its docs and pricing page spell out the use case for a 30-person team so clearly that the engine matched it on thin outside evidence, and it lands last because the vote had so little to count.
D is the one I'd study. It's missing, even though it beats C in any demo. D has no comparison pages, and its homepage talks about a philosophy without naming a category. Its reviews never mention a team size, so it failed at stage one.
All four outcomes come down to published text, which is this piece's whole argument. D shows it best: nobody weighed it and said no, because the pipeline had nowhere to put it.
What doesn't move product rankings, despite the invoices
Plenty of money still goes to work this pipeline can't see, and all of it shares one flaw: it imitates evidence without producing any. A system that reads what the web actually says can tell.
Schema markup by itself barely shifts anything. It helps a parser confirm what the page already says, but it gets no ballot in stage three and can't build a frame in stage one. A listicle on your own site that ranks you first reads as testimony, and engines clearly favor rankings made by someone else.
The press-release burst is the one I'd push back on hardest. It buys a week of extra mentions, which then age past the recency window with nothing to follow them, because stage three rewards voices that keep showing up. Paid one-off placements on random sites fail as well, since they carry no authority and none of the customer language the vote is counting.
Then come the courtesy files, llms.txt among them, and the other technical talismans. None has shown a measured effect at any stage here.
The per-engine wrinkle
Before the playbook, know that engines don't all behave alike, since that decides where you check results. They share the pipeline but not the ingredients. Practitioners who put the same product prompts to every engine find ChatGPT and Claude returning surprisingly similar picks, while Perplexity strays furthest, which fits its heavier live retrieval and denser citations.
Community studies of where citations come from show the split from another angle: each engine has its own appetite, and Reddit and reference sites carry different weight in each.
This is the point we build the whole measurement argument on. You can own the shortlists in one engine and not appear in the next, so product-ranking work that isn't checked engine by engine is being checked on vibes.
15 hours a week manually. Or 15 minutes with RankControl.
Track citations, monitor competitors, and fix content gaps across every AI search engine. Automatically.
The playbook, stage by stage
Start with the frame: write one page that says what the product is and who it's for, in sentences so plain they're almost boring, then add honest comparison pages covering the alternatives your buyers really weigh. Those pages are the frame you want a model to reach for.
Next, put your constraints on crawlable pages in plain text, starting with team sizes and use cases. Integrations, pricing tiers and limits go there too, because an engine can't match an attribute nobody wrote down. Whatever your sales team explains on calls is the spec sheet the engines don't have.
For the vote, keep reviews arriving at a steady pace, with customers saying what they use you for. Take part for real wherever your category gets argued about online, and chase the roundup pages you can see engines citing. Buyers start in the chat window now, so the votes get counted before you know an election happened.
Watching the ordering means running the same buying prompts every time, checked per engine weekly, and logging your position and how you're described, since presence alone tells you little. When a shortlist turns against you, that log points to the stage that broke.
Is a ranking that moves this much worth a system? I'd say the movement is the reason to build one. The same prompt returns a very different mix across runs and engines, so no incumbent has a slot locked up, and you can fight for every stage above this quarter.
In B2B SaaS, the products that win these recommendations have a readable frame and written-down specifics. Their evidence stays fresh, and someone on the team caught the engine that was describing them wrong while rivals were still debating whether AI answers were worth the effort.
AI search traffic grew 835% this year. Is your content ready?
RankControl generates 26 content formats optimized for ChatGPT, Claude, and Perplexity. Published on your domain, matched to your brand.
