A four-step GEO method, and where each step fools you
Everyone knows the order by now: decide which questions matter, measure them, create content and measure the result. What almost nobody tells you is where, in each step, you end up measuring yourself.
Step 1. Decide which questions to measure
The questions to measure come from three sources, and the most reliable one does not come from any search engine: what your customers ask you before they buy.
- What the engine asks itself. The AI performance report in Bing Webmaster Tools publishes the queries Copilot runs before citing you. It is demand observed inside the channel itself.
- What people search on Google. Search Console works, but not raw (the trap is below).
- What your customers ask you. Emails, your contact form, call notes, chat. Nobody else has that data and no bot pollutes it.
The trap: Search Console counts as impressions searches that no person made. On joseredondo.es, between July and September 2026, 28.5% of query impressions came from rows with no clicks at all, and 160 queries looked like prompts from other AI visibility tools running their batteries against Google. A question list pulled from there unfiltered ends up measuring another tool's questions. The full finding.
That is why your customers' questions go first. In the dashboard you paste them as they are and they come out as the question that person would ask ChatGPT, without your brand and without personal data.
Step 2. Measure without biasing the measurement
The bias that matters is not who asks: it is that the same engine does not answer the same way twice.
Asking from your own ChatGPT session is a bad idea: the app may have memory and personalisation on and answer you in particular. API calls carry no history or account memory, so the risk there is a different one, and bigger than it looks:
- The engine is unstable. We ran 269 questions three times in a row. In the 27 cases where the brand appeared at least once, in 12 it did not appear all three times. Finding 02.
- Engines do not resemble each other. Two engines answering the same question share between 4% and 13% of their sources. Measuring one and extrapolating is guessing. Finding 03.
The fix is not a particular tool: it is repeating every question, measuring every engine separately and publishing every percentage with its sample and interval. Details in noise and margin of error.
Step 3. Create content from what you measured
Useful content answers a question you have measured and do not appear in today, on the page where people buy.
Format helps you get extracted: direct answer first, FAQs, lists, a signed author, data with its source. On effectiveness there is only one figure with a study behind it: in Aggarwal et al. (KDD 2024), adding statistics, quotations and references raised visibility by up to 30-40% relative on their benchmark. Any other multiplier, such as "FAQs triple citations", has no published study behind it.
The trap: writing for the wrong question, or writing well on a page that does not sell. Step 4 shows it.
Step 4. Measure the result in four layers, not three
Presence, accuracy, traffic and contact. Almost everyone stops at the first, and the one that decides whether the work pays off is the last.
| Layer | What it tells you | Where it fools you |
|---|---|---|
| Presence | Whether you appear, in what position and how much against competitors | A percentage without its sample and interval cannot be compared with last week's |
| Accuracy | Whether what it says about you is true, and in what tone | Our first accuracy measurement read 99.5%. It was false: 106 of 107 checks could not fail. Done properly, it came out at 64.2% over 53. Finding 04 |
| Traffic | How many visits reach you from an AI assistant | GA4 loses part of it: ChatGPT does not always pass the referrer and it lands in "direct" |
| Contact | Whether those citations land on the pages where people buy, and in enquiries | In the Bing report for joseredondo.es, only 2.6% of 3,665 Copilot citations landed on a sales page. Lots of visibility, little way out |
The fourth layer is what turns a visibility report into a business conversation: which cited pages lead nowhere and which buying questions you are losing.
What this method does not solve It does not turn correlation into causation: outside experiments with a control group, everything measured here is observational. It does not guarantee you appear either: engines change without notice, and in August 2026 Google AI Mode started citing a third of the sources it used to, without anyone touching anything. What it does is tell you at every moment what is happening and why.
Frequently asked questions
What are the steps of a GEO method?
Four, in this order: decide which questions to measure, measure them without bias, create content from what you measured and measure the result. The result is measured in four layers: presence, accuracy, traffic and contact.
Where do the questions to measure AI visibility come from?
From three sources: the queries published in the AI performance report of Bing Webmaster Tools, filtered Search Console, and the real questions your customers ask you before buying. The third is the most reliable, because no bot pollutes it.
Can Search Console queries be used as prompts to measure AI visibility?
Only filtered. On joseredondo.es, between July and September 2026, 28.5% of query impressions came from rows with no clicks and 160 queries looked like prompts from other AI visibility tools. Unfiltered, you end up measuring another tool's questions.
Does always using the same API key bias the measurement?
API calls carry no history or account memory, so the bias that matters is a different one: the same engine does not answer the same way twice. Of 27 cases where a brand appeared at least once in three runs, in 12 it did not appear all three times. That is why every question is repeated and every percentage comes with its interval.
What should you measure in GEO besides presence?
Accuracy (whether what the AI says about your brand is true), the traffic that reaches you from AI assistants, and contact: whether citations land on the pages where people buy. On joseredondo.es only 2.6% of 3,665 Copilot citations landed on a sales page.
Start with step 1
The free audit measures your presence in buying questions for your sector, with its sample and interval, and tells you who appears in your place. No signup, no card.
Run my free audit →Home · Plans & pricing · Methodology · Research · Perplexity · Metrics · How to choose · Use cases · Changelog · Agencies · Alternatives · Diagnosis · Check if AI cites you · Noise and margin of error · Optimizing for agents · X · Product Hunt