Intermediate · Level 2–4 · Tools: LLM APIs, Scraper/monitor, Issue tracker
- Define prompt set
- Query models
- Extract citations/mentions
- Score share of voice
- Alert on regressions
- Create content/entity tasks
More and more people ask AI assistants like ChatGPT, Claude, Gemini or Perplexity for recommendations instead of searching Google. This workflow asks those assistants a fixed set of questions on a schedule, records whether your brand shows up, and flags it when you start disappearing.
A founder asks ChatGPT for the best tool in your category, and your product isn't in the answer. Two competitors are. Nobody knows whether that's new, how often it happens, or whether Gemini says something different. Checking by hand gives you a few anecdotes. This workflow gives you a trend line, which is what you need before deciding whether to do anything about it.
At a glance
- What it does: Track how often AI assistants mention or cite you, and act when that drops.
- Difficulty: Intermediate. Needs a few connected data sources and some care with prompts and review steps. Build a beginner workflow first if you can.
- Automation level: Level 2–4 on the levels of SEO automation. In plain words: anything from a fixed chain of steps (Level 2) up to an AI agent choosing some of its own next steps (Level 4), depending on how far you take it.
- Steps: 6, from "Define prompt set" to "Create content/entity tasks".
What you'll need
Tools
- LLM APIs: direct access for software to AI models like ChatGPT, Claude, Gemini and Perplexity, so you can ask them lots of questions automatically.
- Scraper/monitor: something that records AI answers over time, either a script you build or a monitoring product.
- Issue tracker: wherever your team tracks tasks, like Jira, Asana, Trello or Linear.
What goes in
- A fixed list of questions your customers would ask
- Access to the AI models you care about
- Your brand name and your competitors' names
- Somewhere to store the answers over time
What comes out
- A saved record of every answer
- Your share of mentions and citations over time
- Alerts when that share drops
- Content and entity tasks for your team
How it works, step by step
- Define prompt set. You write 30 to 100 questions (prompts) that real customers might ask an AI assistant, like "best invoicing software for freelancers".
- Query models. On a schedule, it sends each question to each AI model and saves the full answers.
- Extract citations/mentions. It records which brands each answer mentions, and which websites it cites as sources (its citations).
- Score share of voice. It works out your share of voice: the percentage of answers that mention you, compared with your competitors.
- Alert on regressions. When your share drops noticeably and stays down for more than one run, it sends an alert.
- Create content/entity tasks. It opens tasks to improve the pages and facts AI models lean on: your comparison pages, your About page, or your listings on sites that get cited a lot. The facts part is sometimes called entity work, meaning making the basic facts about your brand clear and consistent across the web.
A worked example
Say you track 50 questions across three AI assistants every week. For two months, you appear in roughly 40% of answers. Then, over three weeks, that slides to 25%, mostly on comparison-style questions.
The workflow shows that a competitor's new comparison guide is now cited in many of those answers, and that your own comparison page hasn't been touched in a year. It opens a task to refresh that page and another to check your listings on two review sites the assistants keep citing.
Where AI helps, and where it doesn't
- Reading hundreds of answers and pulling out which brands and sources they mention.
- Grouping questions into themes so you can see where you're weakest.
- Summarising what the most-cited sources have in common.
AI answers change from day to day, even for the exact same question. A single run tells you very little. The workflow should look at trends over weeks and never react to one bad answer.
What to keep human
- A person decides how to respond when a competitor starts showing up more, because that's a strategy question, not a data one.
For more on deciding how much to hand over, see the SEO autonomy dial and what not to automate.
What can go wrong
- AI answers vary a lot from one run to the next, so the numbers can look alarming when nothing has really changed.
- The team overreacts to noise and rewrites pages to chase a single bad week.
Start small
Pick 20 questions and one AI assistant. Run them once a week, paste the answers into a sheet, and mark whether you're mentioned. After a month you'll know whether it's worth automating further, and which questions matter most.