Three Streams GEO Methodology
AI Visibility Scores
Send an AI engine the same unbranded prompts many times, then read three things from those runs: how often your own pages are cited, how often your brand is named next to your competitors, and whether the engine gets your facts right. Each engine is read on its own.
How to choose the competitive set
- Run a few of your monitoring prompts first and pick from the brands that actually appear in the answers: direct competitors, an aspirational competitor and the category leader.
- Use the same set on every engine, so the comparison means the same thing everywhere.
- Keep it for the whole cycle, 9–12 months TSM default; calibrate to your data. Change it only for a real market event, such as an acquisition or a major new entrant, and write down why.
- Never drop a competitor because it is named less often. That is a finding, not a reason to change the set.
- Declare the set, and the prompts you ran, in every report.
Where the counts come from
Your sentinel queries are the prompts customers send that do not name your brand ("which dust extractor suits a small woodworking shop?"). Run each one many times, on different days, from the same declared and network location 📊 Chen & Koudas, arXiv 2610.24081, Oct 2026. Prompts that name your brand are a separate set and never count here.
For every run, note two things. First, whether a page on your own website is among the sources the answer cites. Second, which brands in the set the answer names in its text. Count each once per run, wherever it appears and however often. The Sentinel Query Set Builder helps you write the prompts.
For answer accuracy, check each fact the answer states about your brand (a specification, a price, a policy) against your own records. That check is done by a person.
Keep runs comparable
- Record engine, model and mode for every run. In ChatGPT, instant modes mostly ran one round of searches, while a thinking mode ran several (11.82 searches over 3.20 rounds) 📊 Jiang et al., arXiv 2610.15382, Oct 2026. Runs in different modes are not comparable.
- Leave out pages not yet indexed. A new page that no search crawler has indexed cannot be found by any search, so it scores zero for reasons that have nothing to do with its content. Confirm it is indexed before you judge whether it is cited. TSM default; calibrate to your data.
What matters
- Two counts per run, kept apart: whether one of your own pages is cited, and whether your brand is named.
- Many runs, spread across days: a single run tells you little.
- Each engine on its own, compared with its own earlier periods.
- The same competitive set, prompts, location and mode every period.
- Facts that are wrong: each one is logged and fixed at its source.
- How often you are named, read next to how often each competitor is named, on the same runs.
What matters less than it seems
- Where in the answer you are named. The same brand list appeared in fewer than 1 in 100 repeated answers, and the same order in fewer than 1 in 1,000 🏷 SparkToro/Gumshoe, 27 Jan 2026.
- How many times one answer names you: it still counts once.
- One blended figure for "being in the answer". A citation and a mention come by different routes, so they are counted separately.
- One "AI visibility" score that averages engines.
- A good or bad percentage. No verified benchmark exists; compare each engine with its own earlier periods and with your competitors.
- A vendor's single "visibility" or "share of voice" score. It may weight position or blend engines; go back to the run counts.
Alongside the runs
- ChatGPT referral traffic. "ChatGPT automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs" 🏷 OpenAI, Publishers and developers FAQ, live, checked 2 Nov 2026. Track it in analytics. Treat the count as a floor: the tag went missing for about 7–15 October 2026, which OpenAI acknowledged as a bug, so mark such periods in reports.
- Fan-out queries. The prompts above measure the outcome. The searches the engine runs while answering (fan-out queries) show which passages should match. The Sentinel Query Set Builder explains how to log them.
Sources
- Jiang et al., "Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations", arXiv 2610.15382, 11 Oct 2026 📊 preprint. Instant modes mostly single-round; GPT-5.4 Thinking 11.82 queries over 3.20 rounds.
- Chen & Koudas, "What Drives AI Citations? Content, Retrieval and Model Knowledge in GEO", arXiv 2610.15033, 10 Oct 2026 📊 preprint; API models, not the consumer apps. "Pages are thus cited when searched; brands are also named when remembered."
- Chen & Koudas, "Location, Location, Generation", arXiv 2610.24081, 16 Oct 2026 📊 preprint. Declared and network location change cited sources and recommended brands.
- Schulte et al., arXiv 2604.07585, Apr 2026 📊 preprint. Which brands appear is moderately stable across runs; their order is not.
- Sielinski, "Quantifying Uncertainty in AI Visibility", arXiv 2603.08924, Mar 2026 📊 preprint. Single runs are unreliable; answers are correlated, so real uncertainty is wider.
- Martinez, "Measuring GEO Visibility: Prompt Corpora Define the Answer Market", arXiv 2609.06811, Sep 2026 📊 preprint. The prompt set defines the market measured.
- SparkToro/Gumshoe, 27 Jan 2026 🏷 vendor study. Brand lists and their order rarely repeat.
- OpenAI, "Publishers and developers FAQ", help.openai.com/en/articles/12627856 🏷 live, checked 2 Nov 2026. utm_source=chatgpt.com on referral links.