AI SEO Prompts That Hold Up: Audits, Clustering and Drafting With a Check for Each

A language model answers from two places. The first is parametric memory: what it learned in training, frozen at a knowledge cutoff. The second is live retrieval: pages it fetches while answering, if search is switched on. Neither includes your Search Console data, your rank tracker or your keyword tool.

The cutoff also varies inside one product. As of October 2026, OpenAI's model docs list a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path that ChatGPT also uses. "What ChatGPT knows" is not one fixed thing.

When search is on, the model does its own research. OpenAI's help page says ChatGPT search typically rewrites your query into one or more targeted queries before sending them to search providers. So the pages it reads are chosen by its queries, not yours.

What this means for SEO work is simple. A prompt is good at sorting, comparing, summarising and drafting from material you hand it. It is bad at reporting facts about the web it cannot observe. Every pattern on this page follows one rule: you supply the input, and you check the output. For the wider setup, see using ChatGPT and OpenAI models for SEO work.

Search volume is a measurement from a keyword data provider. Rankings come from a rank tracker or Search Console. A model has access to neither, yet it will usually answer anyway with a number that looks right. That number is a guess shaped like data.

AI answer positions are worse, because no ground truth exists at all. Google's generative AI report in Search Console shows impressions only, with no clicks, no CTR and no position. If Google does not expose a position, no prompt can report one.

Prompts to delete from your library:

• "Give me the monthly search volume for these 50 keywords."
• "What does my site rank for in Google?"
• "What is my position in AI Overviews for this query?"
• "Which keywords have low difficulty?"
• "Estimate how much traffic this page will get."

Replace each one with an export. Pull volumes from your keyword tool, rankings from your tracker, impressions from Search Console. Paste that file into the prompt and ask the model to work on it.

The check is mechanical: if a number appears in the output that was not in your input, delete it. Vendors who sell "AI rank" figures make the same mistake at scale, which is one of the GEO red flags worth knowing.

Be plain with yourself about which part of AI visibility a prompt can touch.

In your control

• The data you feed the model and the checks you run on its output.
• What you publish, including whether key facts are stated plainly on a fetchable page.
• Access: whether AI crawlers are allowed in robots.txt, whether your firewall or CDN lets them through, and whether your content is in the HTML before JavaScript runs.
• Snippet controls such as noindex, nosnippet, max-snippet and data-nosnippet, which keep content out of AI answers when applied, including by accident.

Not in your control

• Whether an engine runs a search for a given question at all.
• Which pages its retrieval step picks and in what order.
• Which retrieved pages the model cites.

No prompt changes the second list. The published GEO studies with effect sizes measured pages already inside the model's context, and the best-designed of them found rewriting tactics ineffective or harmful. That is why there is no proven way to increase AI citations, and why the audit prompts below focus on access.

Audit prompts earn their keep on tedious parsing work. Each pattern below comes with a check that does not depend on the model being right.

1. robots.txt by crawler

Prompt: "Here is my robots.txt. For each of these user agents (GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Meta-WebIndexer, Googlebot, bingbot), state whether it may fetch /, /pricing and /blog. Quote the line that decides each answer."

Check: read every quoted line against the file. Then confirm what each bot does from the operator's docs. OpenAI states its settings are independent: GPTBot governs training, while OAI-SearchBot controls ChatGPT search. A blanket AI block that catches OAI-SearchBot removes you from ChatGPT search answers.

2. What exists before JavaScript runs

Prompt: "Here is the raw HTML of my pricing page, fetched without running JavaScript. List every price, plan name and product claim that appears in visible text. Do not infer anything that is not in this HTML."

Check: fetch the page yourself with JavaScript off and search for each item. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but do not execute them, so anything missing here may be invisible to those crawlers.

3. Accidental suppression

Prompt: "Find every noindex, nosnippet, max-snippet and data-nosnippet in this HTML and these response headers. Say which content each one affects."

Check: search the source for each directive yourself, then ask whoever owns the template whether it was intended.

4. Fact contradictions across your own site

Prompt: "Here are my pricing, about and footer pages as text. List every fact (price, founding year, plan limits, locations) that is stated differently on two pages."

Check: open each page and confirm both versions. Contradictions on your own site are a plausible source of wrong AI answers, and fixing AI answers that get your pricing wrong starts here.

One thing no prompt can check is your firewall. robots.txt can allow a crawler while your CDN returns a 403. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away. Turn any failure into a developer ticket that gets done.

Keyword clustering is where models save the most time, as long as the keywords and numbers come from you.

Prompt: "Here is a CSV of queries from Search Console with clicks and impressions. Group them by search intent. Name each cluster in plain words. Every query must appear in exactly one cluster. Put any query you are unsure about in a cluster called Unsure. Use only the numbers in this file."

Checks:

1. Count the rows. The total in the output must match the total you pasted. Missing or duplicated queries mean the model dropped or invented data.
2. Sum impressions per cluster yourself in a spreadsheet. Totals the model reports should match.
3. Spot-check ten queries at random. Would the same page satisfy each one?
4. For your largest clusters, search a few queries by hand and see whether the results show the same kind of page.

Keep one limit in mind. Clusters built from Search Console describe how people search Google. They do not tell you what ChatGPT searches for, because it rewrites prompts into its own queries. That gap is the main weakness of prompt tracking as a measure of AI visibility.

Drafting prompts go wrong when the model fills gaps with confident inventions. Remove the room to invent.

Prompt: "Write a draft of the pricing FAQ using only the facts in the fact sheet below. Where you need a fact that is not in the sheet, write [NEEDS SOURCE] instead. Do not add statistics, quotes or customer names."

Checks:

1. Highlight every factual sentence in the draft and match it to a line in your fact sheet.
2. Resolve every [NEEDS SOURCE] with a real source or cut the sentence.
3. Search the draft for numbers and names. Any that are not in your inputs come out.

There is one on-page pattern with positive experimental signal: stating a specific fact (a price, a number, a one-line definition) on a page an engine can fetch gives it a sentence it can quote. Ask the model to put those facts in plain sentences.

Skip prompts that ask for an "AI-friendly style", "chunked" sections or long-tail keyword variants. Google's AI optimization guide says you do not need to write in a specific way for generative AI search and that there is no ideal page length.

Volume is the other trap. Google's spam policy lists using generative AI tools to generate many pages without adding value as scaled content abuse. In Otterly's 2026 experiment, two new domains with about 1,000 AI-generated posts each were both removed from Google's index, with no manual action reported in Search Console.

Run prompts in the order that matches the evidence: access first, facts second, content third.

After step 4, check what AI engines say about you today. The free checker on this site runs up to three buyer questions and shows whether you are named, which gives you a snapshot to compare against later. Treat it as a snapshot, since engines decide per question whether to search.

• Asking for data the model does not have. Volumes, difficulty scores, rankings and AI positions. Export them instead.
• Trusting a confident tone. A model's certainty says nothing about accuracy. Every pattern above has a check for this reason.
• Testing logged in with memory on. OpenAI says ChatGPT may use saved memories when rewriting a search query, so your results may not match a buyer's. Test logged out with memory off.
• Treating one answer as the answer. Engines search at their own discretion, so the same question can return a grounded answer one day and a memory-only answer the next. Note whether citations appeared.
• Prompting for schema or llms.txt "to get cited". A model can generate markup, but Google says Search ignores llms.txt and that structured data is not required for generative AI search. Ahrefs' matched test found AI Overview citations fell 4.6% after adding schema. Existing markup does no harm. It just is not a lever.
• Publishing at volume. Prompted pages without added value are what Google's scaled content policy describes.
• Hiding instructions for AI crawlers. Prompt injection and cloaked text exist, and research shows they can sway search agents. They are manipulation, not optimization, and they put your reputation at risk.