How do LLMs choose which brands to recommend? Mostly by searching the web and reading what everyone else says about you. When someone asks ChatGPT, Perplexity, or Google's AI for a recommendation, the model runs live searches, shortlists the brands that rank, then runs follow-up searches to vet each one: reviews, comparisons, the brand's own site. The recommendation goes to whoever survives both passes.
Understand those two passes and the whole game stops looking mystical. Here's each one, and what it demands from you.
Know where the answer comes from: memory vs. retrieval
An LLM has two sources. Parametric memory, which is what it absorbed in training. And retrieval, which is what it fetches from the web right now.
Memory answers timeless questions. But a brand recommendation is never timeless: prices change, companies fold, reputations move. So commercial prompts trigger retrieval. A reliable rule: if the answer depends on what's true today, the model searches. "Best accounting software for a solo restaurant owner" searches. Every "who should I hire" and "what should I buy" searches.
This is the single most important fact in AI visibility, because it means the recommendation layer sits on top of ordinary web search. Whoever wins retrieval feeds the answer. If you've read how to get cited by ChatGPT, this is the same machinery viewed from the brand's side.
Memory still matters at the margins. A brand mentioned consistently across years of web text feels "known" to the model and surfaces more easily in casual conversation. You can't edit training data retroactively. You can make sure everything published about you from today forward says the same clear thing.
Pass one: make the shortlist
The user's conversational question gets rewritten into a few short search queries, the fan-out. "What's a good CRM for a two-person landscaping company?" becomes queries like "best CRM small landscaping business" and "CRM for field service small business." Pages ranking for those queries become the candidate pool, and the brands on those pages become the shortlist.
Note the wrinkle: the shortlist is built from pages, not from brands directly. If the top results are third-party roundups, the shortlisted brands are whoever those roundups include. So pass one has two lanes:
- Your own pages ranking for the fan-out queries. That takes the classic discipline: one page per query, query in the title, meta description, slug, H1, and first sentence.
- Other people's pages that rank: category roundups, directories, review sites. Be present in the credible ones, through legitimate inclusion. If the listicle that ranks #1 for your category doesn't include you, you're invisible in every answer built from it.
Miss both lanes and the process ends here. Nobody vets a brand they never found.
Pass two: survive the background check
Here's the newer behavior, and the one most brands haven't clocked. After the initial fan-out, assistants issue follow-up searches about the specific brands they found: "[brand] reviews," "[brand] pricing," "[brand] vs [competitor]," sometimes searches restricted to the brand's own domain. These are grounding queries, the model going deeper on its candidates before it vouches for one.
It behaves like a diligent, fast researcher. And like any researcher, it trusts corroboration over self-praise. What it's checking:
| What the model looks for | Where it looks | Your move |
|---|---|---|
| What you actually do | Your site | Plain-English entity page: what, who for, where, how much |
| Whether others vouch for you | Review platforms, forums | Real reviews, actively gathered, on third-party surfaces |
| How you stack up | Comparison content | Fair "[you] vs [competitor]" pages of your own, not just theirs |
| Whether your story is consistent | Everywhere at once | Same name, same description, same claims across the web |
Two failure modes end recommendations at this stage:
- Thin evidence. The model searches and finds almost nothing, so it defaults to a better-documented competitor.
- Contradictory evidence. Your site says one thing, directories say another, reviews say a third. Inconsistency reads as risk, and models don't recommend risk.
Watch it happen, then fix what you see
Don't trust our description. Verify the pipeline on your own brand:
- In a desktop browser, open ChatGPT with developer tools on the Network tab, and ask for a recommendation in your category. In the response payloads you can find the actual search queries it issued: fan-out first, brand-specific follow-ups after.
- In Perplexity, ask the same question and read every cited source. That's pass one's winners, in the open. Perplexity SEO is the fastest feedback loop in AI search precisely because it shows its sources.
- Then run the direct probe, in both tools:
I'm considering [your brand] for [use case]. Should I choose them
or a competitor? What are their strengths, weaknesses, and reputation?
Cite your sources.
Save the answers. Every wrong fact is a page you need to fix. Every competitor cited from a roundup you're absent from is an outreach target. Every "limited information available" is a hole in your evidence trail. Re-run monthly and the transcript becomes your progress report.
See the background check in your own data
You don't have to infer any of this from prompts alone. Two free tools show pieces of it happening to your site:
- Bing Webmaster Tools' AI Performance report lists the queries where your content was used to ground AI answers across Copilot, Bing's AI summaries, and partner integrations. Scan that query list for brand-shaped searches: your name plus "reviews," "pricing," "vs." Those are grounding queries hitting your pages. Where the report shows queries your pages barely match, add that language high on the page that's already being retrieved.
- Google Search Console's AI-related query views surface the long, conversational queries reaching you through AI-assisted search. Filter for queries containing "best," "vs," "alternative," and your brand name. Growth there means you're entering more shortlists; the specific phrasings tell you which comparison pages to build next.
Neither tool shows the whole picture. Both show enough to replace guessing with a work queue.
Play both passes, in order
The work sorts cleanly:
For the shortlist (pass one):
- Collect the real fan-out queries behind your category's recommendation prompts
- One answer-first page per query, five placements every time
- Legitimate presence in the roundups and directories that already rank
For the background check (pass two):
- An entity page a machine can summarize in two sentences
- Third-party reviews, real ones, gathered continuously
- Fair comparison pages you control
- One consistent brand description everywhere you appear
- A monthly probe log so drift gets caught early
Do it clean. Fake reviews and manufactured mentions are exactly what these systems are incrementally learning to discount, and a brand caught gaming its own background check doesn't get the benefit of the doubt back. This two-pass structure is why our Reforge Method ends with a hardening phase: pages get you shortlisted, but accumulated, corroborated authority is what makes the recommendation stick.
Start with the probe prompt above. Run it in ChatGPT and Perplexity this week, save what comes back, and fix the first wrong fact at its source, whether that source is your page or someone else's. Next month's run tells you whether it took.