AI SEO Keywords vs Prompts: What AI Search Actually Uses

A keyword is what someone types into a search box. Keyword tools measure it as monthly search volume. A prompt is what someone types into ChatGPT, Gemini, Claude or Perplexity. A retrieval query is what the engine sends to a search index to find pages for its answer.

Only the retrieval query decides which pages the model reads. On most platforms, publishers can't see it.

Say you run a payroll software company. A buyer asks ChatGPT which payroll tool suits a 20-person agency that pays contractors in three countries. That sentence is not what gets searched. The engine may break it into shorter, targeted searches about contractor payroll, multi-country compliance and small-agency pricing. Your page competes on those hidden searches, not on the buyer's original wording.

This is why classic AI keyword research still helps with topics but can't tell you what an engine actually searched. Keep both ideas in your head: keywords size demand, prompts represent intent, and retrieval queries do the real work out of sight.

Each major surface documents some version of query rewriting. None of them, except Microsoft, shows site owners the result.

Two more facts make this harder. First, retrieval is optional. As of October 2026, ChatGPT decides per question whether to search, and Google AI Overviews often don't trigger at all. Second, an answer given without search comes from training data. OpenAI's model docs list a February 16, 2026 knowledge cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. No query of any kind reaches your site in that case.

What it means: the same prompt can produce different searches, or no search, from one run to the next.

Prompt tracking tools send a fixed list of prompts to each engine and record who gets named. That's useful. It is not a measurement of the queries real buyers trigger.

The gap comes from four places:

• Rewriting. The engine searches for its own reformulation, not your prompt.
• Fan-out. One prompt can become several searches across subtopics.
• Discretion. The engine may not search at all and answer from memory instead.
• Context. Logged-in users bring memory and custom instructions that change the rewrite.

Nobody outside the engine sees all of this. Tools can't see the prompts people actually type either, a point covered in whether AI visibility tools can see real AI prompts.

The next move is to treat your prompt set as a sample of buyer intent. Write prompts that cover the questions buyers really ask, at the stages they ask them. The prompt research method walks through how to build that set. Then report results as "named in X of these sampled answers on this engine", never as a rank.

Bing Webmaster Tools added an AI Performance report in public preview on February 10, 2026, and expanded it in June 2026. It includes grounding queries, which Microsoft describes as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers". The report covers Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. It also shows citations, cited pages, intents, topics and citation share.

This is the only place any engine shows publishers the actual retrieval input. Everything else in the market is a prompt someone chose.

Carry three caveats:

• It's a sample. Microsoft says the grounding query data represents a sample of overall citation activity.
• It's Microsoft only. It tells you nothing direct about ChatGPT, Gemini or Google AI Overviews. The Bing index vs Google index comparison explains which answers draw on which index.
• It's UI only. When this was last verified in August 2026, the report had no API, so you export it by hand.

Microsoft also frames citation share as an observational metric, not a ranking or a competitive scoreboard. Read it that way.

The next move: verify your site in Bing Webmaster Tools, export grounding queries each month, and compare them with the prompts you track. Where they diverge, the grounding queries are closer to reality.

Not in your control

• How an engine rewrites or fans out a prompt.
• Whether it searches for a given question.
• Where your page lands in the model's list of retrieved documents.
• What a model learned before its training cutoff.

In your control

• Access. AI crawlers have to reach your pages. robots.txt rules are per bot: OAI-SearchBot governs ChatGPT search, while GPTBot governs training only. A firewall or CDN can still block a crawler that robots.txt allows.
• Rendering. Most AI crawlers other than Googlebot don't run JavaScript, so content that appears only after scripts run may be invisible to them.
• Index eligibility. Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode, with no extra technical requirements.
• Topical coverage. A critical survey of 45 GEO studies (arXiv, July 2026) found topical relevance and context position the most reproducible levers. Those studies measured pages already inside the model's context, so read the gap between lab GEO studies and live engines before treating that as a promise.
• Specific facts. A page that states a price, a number or a one-line definition gives an engine a sentence it can quote.

None of this guarantees a citation. It removes the reasons you can't be cited.

1. Check access first. Confirm robots.txt allows the search bots you care about (OAI-SearchBot, PerplexityBot, Claude-SearchBot) and that your firewall returns them a normal page, not a 403.
2. Verify both webmaster tools. Set up Bing Webmaster Tools and Google Search Console. One healthy index tells you nothing about the other.
3. Pull grounding queries. Export them from Bing's AI Performance report and group them by topic and intent.
4. Build a prompt set. Combine real buyer questions with the themes in your grounding queries. Use keyword volume to size topics, not to pick exact wording. A free way to find the questions buyers ask AI covers the no-cost version.
5. Map topics to pages. For each topic, make sure a fetchable page exists and states the key facts in the HTML.
6. Test engine by engine. The free checker on this site shows whether AI engines name you for up to three buyer questions, as a one-time snapshot. Read each engine's result separately.
7. Repeat monthly. Re-export grounding queries, compare them with your prompt set, and update the set where real queries have moved.

• Stuffing long-tail variants. Google's AI optimization guide says you don't have to worry about capturing every variation of how someone might search. In the original GEO lab study, keyword stuffing lowered the authors' visibility metric to 17.8 against a 19.5 baseline.
• Treating generated keyword lists as demand. An AI keyword generator produces ideas, not evidence that anyone searches for them.
• Reporting an "AI rank". Google's generative AI report in Search Console shows impressions only, with no position data to check a rank against.
• Blending every engine into one score. Engines use different indexes and rewrite differently. In Otterly's 2026 test of two AI-generated sites, Microsoft Copilot cited one site thousands of times while ChatGPT almost never did.
• Assuming the grounding queries are the whole picture. They are a sample from Microsoft surfaces. Use them to correct your prompt set, not to replace it.
• Rewriting pages in an "AI-friendly" style. Google says you don't need to write in a specific way for generative AI search. Spend that effort on access and on stating facts clearly.