Find the questions AI assistants answer without naming your brand, and plan content to fix it
Write the questions your buyers ask, run each one many times across several AI assistants, and record how often your brand and your competitors appear and which pages get cited. Gaps are questions where competitors show up and you do not. Plan one useful page or fix per gap. A large SparkToro study found AI brand lists change from run to run, so one check proves nothing. Measure presence over many runs, not rank.
You get a table of buyer questions with your mention rate, competitors' mention rates and cited sources per assistant, plus a short content plan for the biggest gaps that you approve.
- Difficulty
- Intermediate
- Time
- Half a day for the first run, then a monthly repeat
- Sources
- 8 · 8 independent
- Updated
- 2026-10-13
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
Help me find questions where AI assistants mention my competitors but not my brand. My brand: [brand], website [domain], which sells [offer] to [buyer]. Competitors: [3-6 names and domains]. Buyer questions: [paste 15-40 real questions from sales calls, support tickets, search console queries or forums. If I have none, draft 30 and STOP so I can edit them.] Assistants to check: [ChatGPT with search, Perplexity, Gemini, Claude with web search, Google AI Overviews]. Do this: 1. Run each question 5 or more times on each assistant, in fresh chats, logged out or in a clean profile where possible. Record the date, assistant, run number, whether my brand is named, which competitors are named, and every cited URL. 2. Build a table per question: my mention rate (runs where I appear / total runs), each competitor's mention rate, and the domains cited most. 3. Mark a gap where a competitor appears in at least 40% of runs and I appear in fewer than 10%. 4. For each gap, open the top cited pages and say in two sentences what they cover that my site does not. Check my site for an existing page on the topic. 5. Propose at most 5 actions, ranked by how many gap questions each one covers. An action can be a new page, an update to an existing page, a fix to a wrong fact about us, or a listing on a third-party site the assistants cite. Rules: - Report rates, not single answers. Do not call a question a gap from one run. - Note any answer that states something wrong about my brand, with the exact text. - Never suggest hidden text, prompt injection, fake reviews or mass-produced pages. - STOP after the table and the plan. Do not write or publish content until I approve.
Human checkpoints: the agent stops and asks you
- Approve the list of buyer questions before the agent runs them.
- Approve the content plan before anything is written.
- Review and publish each page yourself.
The audit itself is safe. The risk is acting on noise, or publishing many thin AI pages to chase answers, which Google's spam policies call scaled content abuse. You approve the plan and every page before it is published.
What you need
- Real buyer questionsdata
No assistant shares prompt volume. HN commenters who track this say tracked prompts are guesses, so start from questions buyers actually asked you.
- Access to several AI assistantsaccount
Consumer apps for spot checks, API keys if you script the runs.
- A spreadsheet or results filedata
Steps
Collect buyer questions
Pull questions from sales calls, support tickets, Search Console and community threads. Keep the buyer's words.
Check: You have 15-40 questions that a buyer would really type.
Run each question many times
Run each question at least five times per assistant in fresh, logged-out sessions. SparkToro's study found brand lists almost never repeat exactly between runs, and commenters report location and account history change answers too.
Check: Every question has several runs per assistant with the date recorded.
Compute mention rates and find gaps
Let the agent compute your mention rate and each competitor's rate, and flag gaps. Ignore position in the list.
Check: The table shows rates per question and per assistant.
Read the cited pages
For each gap, open the pages the assistants cite. Look for what they explain that you do not, and for third-party sites (directories, reviews, Wikipedia-style profiles) that name competitors but not you.
Check: Each gap has a short note on why competitors get cited.
Approve a small plan
Pick a few actions that cover the most gaps. Prefer fixing wrong facts and improving pages people already use. Google says no special optimisation is needed for its AI features beyond normal search best practice.
Check: You have at most five approved actions, each tied to named gap questions.
Re-measure after a month
Run the same questions the same way after the changes are live. Compare rates, and expect some movement from normal drift.
Check: You have before and after rates for the same questions.
What goes wrong
Someone checks one answer, sees no brand mention and declares a gap.
Fix: Run each question many times and use the mention rate. SparkToro suggests 60-100 runs per prompt for tracking, and one HN poster's single-check claim could not be reproduced by commenters.
Answers change week to week, so a gain or a loss may be drift.
Fix: Re-measure with the same method and compare against competitors' rates in the same runs, not against last month alone.
Personalisation, location and account history change the answers.
Fix: Use fresh, logged-out sessions or APIs, and record the setup with the results.
The assistant confuses your brand with another entity or repeats outdated facts.
Fix: Record the exact wrong text. Make sure your site and your third-party profiles state the facts plainly and link to each other. One agency found ChatGPT and Google's AI Overview resolved the same brand in different ways.
The plan turns into many thin pages written to chase AI answers.
Fix: Publish fewer, useful pages. Google's spam policies name AI-generated pages at scale without value as scaled content abuse.
Cited links in answers are broken or point to copies of the real source.
Fix: Open every cited URL before you draw conclusions. The Tow Center found AI search tools often cited wrong or broken links.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- Blog2026-01-28AIs are highly inconsistent when recommending brands or products
SparkToro and Gumshoe had volunteers run the same prompts thousands of times on ChatGPT, Claude and Google AI. The same brand list almost never repeated between runs. The authors advise tracking how often a brand appears, not its rank, and list API versus user-facing output as an open question.
Rand Fishkin · sparktoro.com - Hacker News2026-03-12How do you measure your brand's presence in LLM outputs?
A marketing director says a year of manual prompting across ChatGPT, Perplexity and Claude, tracked in spreadsheets, showed his brand sometimes absent or described in vague or outdated terms. He found no standard tooling.
marcosviladomiu · news.ycombinator.com - Hacker News2026-02-05· ChatGPTAsk HN, how does ChatGPT decide which websites to recommend?
The poster and commenters report the same prompt giving different recommendations day to day, a product recommended for a week and then gone, and no insight into what users actually ask.
nworley · news.ycombinator.com - Hacker News2026-03-20Launch HN, Sitefire
A startup runs synthetic prompts daily on several AI engines. Commenters raise that location, IP and account history change answers and add noise that is hard to control. The thread gives no answer to that problem.
vincko · news.ycombinator.com - Blog2026-09-23· ChatGPTGoogle found the pages, ChatGPT found the entity
An agency ran 20 logged-out chats and says ChatGPT identified a client brand each time via a third-party profile, while Google's AI Overview cited unrelated entities with a similar name. It recommends linking the official site to external profiles.
Stefan Petschinka · richresults.ai - Hacker News2025-06-06Get your dev tool mentioned by ChatGPT and Gemini
The poster says a ChatGPT query left out his product and built a tool around hidden prompts. A commenter could not reproduce his results, and another called hidden prompts unethical.
vinodvarma24 · news.ycombinator.com - Hacker News2026-06-24Show HN, estimating what users ask ChatGPT about your company
The builder estimates likely prompts because the AI companies share no query data. In the thread he agrees answers are random enough that each query should be run several times per cycle.
biduskamil · news.ycombinator.com - Hacker News2026-02-02· Custom agentIn my Google Calendar, every event is an AI task
The author runs a recurring citation gap analysis with Claude and search tools as one of several scheduled workflows. No results or problems are reported.
ihmissuti · news.ycombinator.com
- Google Search Central, AI features and your websitechecked 2026-10-13
- Google Search spam policies (scaled content abuse)checked 2026-10-13
- Tow Center, AI search engines are bad at citing newschecked 2026-10-13
- Aggarwal et al., GEO, Generative Engine Optimizationchecked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…