Grounding queries are the follow-up searches an AI assistant runs after its first round of research: narrower, more pointed queries that dig into the specific brands it just found, so it can verify them before recommending anyone. The first wave of searches builds a shortlist. Grounding queries run the background checks. If your business survives the shortlist but fails the background check, you were considered and then quietly dropped. You'll never know it happened.
That second round is where AI recommendations are decided. Here's how it works and how to win it.
How do grounding queries fit into an AI search?
An AI answer is built in stages, and search happens at least twice.
Stage one is query fan-out: the assistant splits the user's question into several broad searches ("best accounting software for freelancers," "accounting software pricing comparison") and collects candidates.
Stage two is grounding. Now the assistant has names. So it searches again, and this time the queries contain those names: a specific brand plus "reviews," a specific brand plus "pricing," a brand's name next to a competitor's. Assistants will even search within a single company's site to pull specifics, the way a careful human would use a site: search. The model learned something from round one and is following up. It's the same loop a diligent buyer runs, compressed into seconds.
The stages ask different questions of your business:
| Fan-out queries | Grounding queries | |
|---|---|---|
| When | First, from the user's question | After, from what the fan-out found |
| Shape | Broad category searches | Narrow brand-specific searches |
| Question being asked | "Who's in this market?" | "Is this one any good?" |
| What it checks | Your category pages and rankings | Your reviews, pricing, comparisons, reputation |
| What losing looks like | You're never considered | You're considered, vetted, and dropped |
Most businesses doing GEO work obsess over stage one. Stage two is where the recommendation is won.
Why does the second stage exist at all? Because assistants get punished for bad recommendations. A model that names a company with terrible reviews, or invents a plan that doesn't exist, produces an answer the user can falsify in one click. So the systems check before they commit. Grounding is the checking, and that framing should change how you write every page an assistant might vet: you're not persuading a reader, you're supplying evidence to a fact-checker.
What do grounding queries look like?
Predictable, which is good news. They're the questions any skeptical buyer asks once they have a shortlist:
- "[brand] reviews"
- "[brand] pricing"
- "[brand] vs [competitor]"
- "is [brand] legit"
- "[brand] alternatives"
- "[brand] for [specific use case]"
Run the shortlist test yourself: ask ChatGPT or Perplexity to recommend a provider in your category, then watch what it cites when it discusses each named brand. Review sites. Comparison pages. Pricing pages. Product and service pages. The pattern holds across engines: assistants ground their recommendations in evaluation content, not in brand slogans. Perplexity is the most transparent about it, since it cites nearly every claim. We dig into its behavior in how to get recommended by Perplexity.
Where can you see real grounding queries?
Bing Webmaster Tools ships a free AI Performance report covering Copilot and Bing's AI summaries. Inside it: the grounding queries that led AI systems to your content. Actual search strings, not reconstructions. The setup and the full read of that report are in our guide to the Bing Webmaster Tools AI Performance report.
The highest-value thing to do with that report is hunt for language mismatches. Compare the grounding queries against the pages they landed on. When AI searches "does [brand] offer month-to-month plans" and your pricing page only says "flexible engagement models," you've found a gap: the machine asked a question your page almost answers.
The fix is usually not a new page. Add the query's language, high up, to the page you already have. A retrieval system rewards literal matches, and one strong page that speaks the query's language beats three thin pages that don't. More pages aren't always better. Better pages are better.
Bing's data is a sample of one ecosystem. But the patterns it exposes generalize, because how machines phrase questions about businesses like yours doesn't change much between engines. Fix the mismatch once and you've fixed it everywhere that searches similarly. For the measurement side across GA4 and Search Console, see how to measure AI search traffic.
How do you win the grounding round?
Grounding queries are a vetting process. You win vetting the way trustworthy businesses always have: by having good answers ready where the checker looks. Two surfaces matter.
On your site, make verification effortless:
- A real pricing page. "Contact us for pricing" is a failed grounding query. Publish numbers, or at least ranges and packages. We publish ours for exactly this reason.
- Fair comparison pages. "You vs. competitor" and "alternatives" pages answer the assistant's comparison queries in your own words instead of leaving that answer entirely to third parties. Fair means fair: name real trade-offs, because a page that admits weaknesses reads as evidence, and a page that admits none reads as an ad.
- Entity FAQs. Plain answers to "who is [brand]," "what does [brand] do," "is [brand] legitimate," on your site, in query-shaped language.
- Specific service and product pages. Grounding often targets a use case: "[brand] for dental practices." A dedicated page wins that query; a generic homepage doesn't.
Off your site, make the story consistent:
- Reviews where machines look: Google, industry directories, review platforms. Volume helps; recency and specificity help more.
- Third-party mentions that describe you accurately: press, podcasts, directory profiles with real descriptions. Every consistent description is another document confirming the same facts.
- Consistency above all: if your site says one thing and the third-party record says another, or says nothing, the assistant has no way to resolve the conflict in your favor. It doesn't adjudicate. It moves on.
One line on ethics, because this corner of the industry has a dirty version of everything above: fake reviews, laundered press, astroturfed "review summary" sites. Skip all of it. You'd be feeding fabricated evidence to systems built to cross-check evidence, on behalf of a brand that has to survive the scrutiny. Real reviews syndicated widely beat fake ones, and they don't detonate later. The full picture of what models weigh when they choose is in how LLMs pick brands.
What should you do this week?
- Ask ChatGPT and Perplexity for recommendations in your category. Note who gets named and what sources get cited when each brand is vetted.
- Ask them directly about your brand ("is [brand] good," "[brand] reviews") and read what comes back. That's your grounding profile, and it's what a prospect's assistant sees.
- Set up Bing Webmaster Tools and open the AI Performance report. Pull the grounding queries.
- Fix language mismatches on existing pages before writing anything new.
- Close the structural gaps: pricing page, comparison pages, entity FAQ.
- Audit the off-site record for consistency, and start filling the holes with real evidence.
This is the sequence we run for clients: audit the grounding profile, rebuild the pages that fail it, then harden the off-site evidence so citations stick. It's the Reforge Method in miniature.
Start with step one before you close this tab. Ten minutes with an assistant gets you the shortlist in your category and the sources it trusted for each name. Write both down. That's your baseline, and every fix on the list gets measured against it.