How to Automate Keyword Research for GEO: A Five-Step Workflow

Last updated: 2026-10-12

How can you automate keyword research for GEO?

You automate keyword research for GEO in five steps: pull the queries you already appear for, add the phrases AI systems used to retrieve your pages, send your customers' questions to the engines, store what they search and cite across repeated runs, and rank the gaps. Steps 1, 3 and 4 can run on a schedule; step 2 starts as a manual read of the report, and step 5 ends with a person's judgement.

1. Pull the queries your pages already appear for. The Google Search Console API returns them with clicks, impressions and position, page by page.
2. Add the phrases AI systems used to retrieve your pages. Bing Webmaster Tools reports them as grounding queries in its AI Performance report.
3. Send your customers' questions to the engines. Call the OpenAI, Claude and Gemini APIs with web search enabled, one prompt at a time. Start from the Search Console queries phrased as questions.
4. Store the searches each model ran and the pages it cited. Run every prompt several times, because the same prompt rarely returns the same answer.
5. Rank the gaps. Compare the cited pages with your own and decide which pages to write and which to fix.

The output is different from a classic keyword list. A keyword tool gives you a phrase and a monthly volume. This workflow gives you a question, the searches a model issued to answer it, and the pages that earned the citation. That is the material a content plan for AI answers needs. It sits next to the search data you already use for SEO (search engine optimization), and it feeds the work GEO and SEO share.

Which data can a script collect before it asks any AI engine?

Three first-party inputs come before any prompt: your Search Console queries, which a script can pull, Bing's grounding queries, and your own server logs. The table shows what each one adds to the workflow and the limit to plan around.

Search Console already contains part of the AI picture. Google states that sites appearing in AI Overviews and AI Mode are counted in the "Web" search type of the Performance report, and the Google pages we read in October 2026 describe no separate filter for them. On that evidence, the Search Analytics query method returns those impressions together with classic results. Google's help page adds that every link inside an AI Overview is assigned the same position.

Bing is the only source here that names AI retrieval directly. Microsoft defines grounding queries as the key phrases the AI used when retrieving content that was referenced in an answer. The announcement describes a report, so plan to read this data by hand until you have confirmed programmatic access. For the logs, OpenAI's crawler documentation lists the names its crawlers use to identify themselves, which is what you filter on. Use the fetched pages as a check alongside step 1.

Which workflows automate GEO keyword research?

Three workflows automate GEO keyword research: a scheduled script that calls the engines' APIs, an AI agent connected to your data sources, and a ready-made keyword research skill.

The idea behind all three is query fan-out. Google's documentation says AI Overviews and AI Mode may use it, issuing multiple related searches across subtopics and data sources to develop a response. Searches of this kind are the keywords of GEO. Three official APIs return the ones their own model ran for a prompt, which is the closest observable equivalent. The table compares what each response contains.

Perplexity's documentation says support for Sonar Chat Completions ended on September 27, 2026, so a tutorial written for that endpoint is out of date.

A scheduled script is the plainest of the three: it loops over your prompts, calls each API and writes queries and sources to a table, and it needs someone who can write and maintain code. An agent does the same work inside one conversation: an AI assistant reads your data through a connector (the standard is called MCP, Model Context Protocol) and runs the steps when you ask. A ready-made skill is a packaged set of instructions for such an assistant, and the ones that rank for this topic cover discovery, clustering and scoring. Before you adopt one, check whether it stores the searches the engines ran and the pages they cited.

How many times should each prompt run?

The only public rule of thumb is at least 60 to 100 runs per prompt, and it comes from the author of a single study, who leaves the statistically sound figure open. A single run tells you little about what an engine usually answers. The best public evidence comes from research by SparkToro and Gumshoe.ai, published in January 2026. In November and December 2025, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI Overviews a combined 2,961 times. The engines returned the same list of brands less than 1 time in 100, and the same list in the same order less than 1 time in 1,000.

Rand Fishkin, who wrote the study, suggests averaging the results of those runs. Gumshoe.ai sells AI tracking tools, so read the recommendation with that in mind.

The searches behind the answers move too. A Surfer study of 173,902 URLs, on a page last updated in September 2026, found that only around 27% of fan-out queries stayed consistent across repeated searches for the same prompt. Surfer sells an SEO tool and generated its fan-out queries with Gemini, so the figure indicates a pattern and is not a measurement of Google's own system.

Repetition is what turns this into a budget line. Anthropic's documentation, read in October 2026, prices web search on the Claude API at $10 per 1,000 searches, plus the usual charge for the text the model reads and writes. As an example, 100 prompts run 60 times at 3 searches each make 18,000 searches, or $180 before that charge. Decide the number of prompts first, then the number of runs.

How do you turn captured queries into a content plan?

You turn captured queries into a content plan by grouping them by topic and checking, for each group, whether one of your pages is among the cited sources. Count how often each search recurs across runs, then list the groups where competitors' pages are cited and yours are missing. Those are the pages to write or fix first.

The Surfer study suggests why the fan-out queries deserve the attention. Across the top 10 pages for 10,000 keywords, pages were 161% more likely to be cited in an AI Overview when they also ranked for fan-out queries. In the same sample, 67.82% of AI Overview citations did not rank in the top 10 at all. The author stresses that the data shows correlation and not causation. Read it as a reason to cover the related searches, and not as a promise of citations.

Do not wait for a volume number to rank the list. In April 2026 Semrush published an analysis of more than 1 billion lines of U.S. browsing records from October 2024 to February 2026. For most of that period, between 65% and 85% of ChatGPT prompts could not be matched to any traditional search keyword. Semrush sells SEO tools and the data covers ChatGPT in one country. Rank by what you can observe: recurrence across runs, your Search Console impressions for the same topic, and who is cited today.

This is also where we start. Every engagement at Oltre AI opens with an audit of how the brand appears in classic search and whether AI engines cite it, and the gaps come out per topic and per engine. That baseline is what later results are measured against. The same gap list feeds a brand mention and source gap analysis and gives AI citation tracking its starting point.

Which parts of GEO keyword research still need a person?

A person still chooses the prompts, judges the gaps and accepts what the data cannot say. The script cannot know which questions your customers ask before they buy. Start from sales calls, support tickets and the Search Console queries phrased as questions, then let the workflow expand them.

The data has limits a person must keep in view. We found no official source in which an AI company publishes how often a prompt is asked, and our search was not exhaustive. Plan the workflow to work without a volume figure. OpenAI's documentation says the search queries are usually, but not always, included in a response. None of the documentation passages we read says an API response matches what a person sees in the app. Treat the captured searches as a sample of how the models behave.

The last judgement is about the pages themselves. Google states there are no additional requirements and no special optimizations needed to appear in AI Overviews or AI Mode. Keyword research for GEO does not replace the rest of the organic work. It tells you which questions to answer, and the pages still need to be crawlable, indexed and worth citing. Measuring whether they are belongs to the KPIs you report for AI search visibility. Repeating the keywords you found will not get you there: the paper that introduced GEO, accepted to KDD 2024, tested content changes on a benchmark of 10,000 queries and found that keyword stuffing offered little to no improvement.

The workflow gives you questions, the searches the models ran and the pages they cited, and a person decides what to write from them.