ChatGPT doesn’t pick brands at random. Neither does Gemini, Claude, or Google’s AI Overviews.
I look at three doors into an AI answer: memory, retrieval, and selection. Each calls for different work. Before you commission another page, find the gap you need to fix.
Here’s what I look for, and what I’d fix first.
The 30-second version
| Door | What happens | What opens it | What your team calls it |
|---|---|---|---|
| 1. Memory | The model often seems to have a shortlist of brands before it searches. | A clear identity and useful coverage in the places buyers trust. | Brand, PR, earned media |
| 2. Retrieval | The model fires off hidden searches to Google, Bing, or its own index and merges the results. | Ranking for those hidden searches, not just your head keyword. | SEO |
| 3. Selection | The product selects information for an answer. | Answer-first pages, real facts, real prices, visible dates. | AEO, content design |
GEO is all three doors.
That’s why “GEO is just SEO” and “SEO is dead” are both wrong. SEO supports retrieval. Brand knowledge and useful content deserve a plan too.
Door 1: Memory — the shortlist before the search
Here’s the part almost nobody talks about.
Before an AI tool searches, it often already seems to know which brands it wants to check. People who watch ChatGPT’s event stream report that it names brands first, then sends searches straight to those brands’ websites (Suganthan Mohanadasan).
A second clue: one study of about 50 prompts compared GPT-5.4’s citations with the Google and Bing results for the same prompts. 75% of the domains ChatGPT cited weren’t in those search results (Writesonic, a tool vendor). The authors connect that gap with training knowledge. My take is simpler: your Google rankings aren’t the whole picture.
So what?
A searchable page isn’t a whole brand strategy. You also need to be part of the conversation around the problem you solve.
How I work on memory:
- Mentions next to the words buyers use for your category. This is my read of the evidence, not a law. Models seem to connect brands with the words that sit near them. A trade article that explains your fit for multi-state payroll gives a buyer more to use than a generic award write-up.
- A steady flow of new mentions. New coverage helps twice: tools that search live can find it now, and future model versions may learn from it later. Don’t expect this month’s press hit to change what a model already “knows” this month.
- Sources beyond your own site. AirOps, a content software vendor, reported that 85% of brand mentions in its October 2025 analysis came from external domains.
- The same description of you everywhere. Keep the category, offer, and important facts consistent across legitimate profiles.
My take: earned reach belongs in the search plan. Here is how I map the sources that matter → The useful story someone else tells about you can matter alongside the useful information you publish yourself.
Door 2: Retrieval — the searches you never see
AI search tools can turn your question into related searches called fan-out queries. Google’s own guidance describes this process. The words buyers type can be the start of a much larger research task.
The hidden searches give you another way to think about coverage:
- They’re long. ChatGPT’s average fan-out went from 6 words to a peak of 16 in December 2025, then settled around 12 (Peec AI, a visibility software vendor).
- They serve a decision. A platform buyer may need information about security, integrations, migration, and pricing. Cover the decisions behind the broad phrase.
Rank still matters. A lot. One analyst processed more than 2 million citations and found roughly a 46–48% chance of a citation at Google position 1, about 37% at position 2, and about 19–20% at position 10 (Brie Moreau, via the Common Crawl blog. He calls the data preliminary).
But rank is a ticket, not a trophy. Read those numbers again: even at #1, you miss the citation about half the time. Retrieval gets you into the room. It doesn’t hand you the microphone.
How I work on retrieval:
- Find the hidden searches behind your most valuable prompts. Then cover those buying questions on strong, coherent pages: one page can answer a whole cluster of related questions. Don’t build a thin page for every query. Google warns against scaled content made primarily to manipulate rankings.
- Rank across the whole topic, not one keyword. ChatGPT merges many result lists, so one broad keyword leaves too much of the buying journey unexamined.
- Protect your Google rankings like revenue. Check access, indexing, canonical URLs, and your important pages before buying another layer of software.
Door 3: Selection — the passage that wins
Don’t count on the model reading your whole page.
Consider an enterprise platform that promises “seamless migration.” A buyer needs to know what moves, who does the work, which systems connect, and what can delay the process. The slogan answers none of that.
How I work on selection:
- Lead with the answer. On the page, and at the top of every section.
- Make every section stand on its own. A block that starts with “As mentioned above…” is useless once it’s cut loose.
- Put real facts on your own domain. Prices. Specs. Service areas. What your product isn’t good for. Give buyers and search products a reliable place to check the facts.
- Keep useful facts current. Show a date when it helps. Do not change a date to simulate freshness.
- Say what’s false. A brand facts page can correct the wrong claims models keep repeating.
Where would I look first?
Start with what you can observe. Then choose the investigation that could change your next decision.
| What you see | Investigate | First move |
|---|---|---|
| The AI names your competitors without searching. Never you. | Memory | Earn recent third-party mentions next to your category words. |
| Your pages don’t show up in the sources the AI opened. | Retrieval | Check access and cover related questions on strong, coherent pages. |
| Your page got opened, but a competitor’s passage got used. | Selection | Rewrite that passage: answer first, real facts, clear date. |
| The AI describes you — with wrong or stale facts. | Memory + selection | Fix the facts at the source and on a brand facts page. |
| You show up some days and not others. | The measurement | Repeat the checks before changing the plan. |
Measure a pattern, not a screenshot
AI answers change every time you ask.
In one study, a hospital showed up in 69 of 71 ChatGPT answers — but was the first mention in only 25 (SparkToro).
That screenshot your agency sent? It records one answer. It can’t tell you how often the brand appears. Keep the question, product, date, and visible sources, then repeat the check.
How I measure it:
- Build a prompt set in your buyers’ own words, grouped by intent.
- Run each prompt many times, per engine. Never blend engines or models.
- Report the share of valid runs that name you, with the counts behind it. Keep failed checks separate.
- Ask every new lead how they found you — open text, not a dropdown. If they say ChatGPT, ask for the prompt.
- Watch branded search and direct traffic next to AI referral clicks. A lot of AI influence never shows up as a click.
What I do first in a 90-day engagement
- Weeks 1–2: The Three-Door Audit. I run your buyers’ prompts and choose what to investigate for each one.
- Weeks 3–6: The first useful change. Start with work the team can actually approve and publish.
- Weeks 4–12: Retrieval and memory. Find the hidden searches, fix the pages that should rank for them, and start earning mentions on the sites AI cites for your category.
- Every month: re-measure. Each new default model can reshuffle results, like a Google core update.
Request your free manual AI visibility snapshot →
Questions people ask
Is GEO just SEO? I use GEO for the wider work of helping a brand appear accurately in generative answers. SEO supports retrieval. Brand knowledge and useful content also need attention. Call the work what your team calls it; plan for all three.
Can anyone guarantee a ChatGPT recommendation? No. AI answers change run to run and model to model. A useful plan defines the work, records what happened, and measures the result honestly.
Does llms.txt help? Google’s current AI search guidance does not require a special AI file or special schema. Follow the requirements of the product you are targeting. A file cannot replace useful information.
Should we publish “best X” lists that rank our own product #1? Use a transparent comparison when it helps the buyer. Disclose your interest, explain the criteria, and show when someone should choose another option. A large collection of self-serving rankings needs a better reason to exist.
A note on the evidence
Most numbers about AI search come from companies that sell AI tracking tools or AI search services. Three of the most active researchers — Malte Landwehr, Tomek Rudzki, and Metehan Yeşilyurt — work for the same company, Peec AI, so I count their data as one source. Plenty of findings come from a single site, a single client, or a single model version. ChatGPT numbers can expire in weeks.
Search-result overlap cannot prove a training-memory cause. Visible sources do not reveal every retrieval step. An opened page cannot prove why a product selected its final answer. Repeated checks from one account are not a random sample of every buyer.
The three-door framework is my synthesis. The useful test is whether it helps us choose better work. I re-test important findings before using them to guide a client’s budget.
Explore the source-checked AI search statistics →
Sources
- AirOps: The influence of off-site signals in AI search. Content software vendor; October 17, 2025.
- Suganthan Mohanadasan: ChatGPT decides before it searches. Event-stream analysis.
- Writesonic: GPT-5.4 citation study. Software vendor; 50 prompts.
- Peec AI: 20 million search fan-out queries. Visibility software vendor; historical query analysis.
- Common Crawl: Web graph data and AI ranking research. January 19, 2026; preliminary work involving more than two million citations.
- SparkToro: Consistency of AI brand recommendations. Research with Gumshoe.
- Google: Current AI search guidance.
- Google: Spam policies.
Sources reviewed November 14, 2026.
Try this with your team.
- Choose one question a real buyer asks before building a shortlist.
- Repeat it in the consumer products that matter to your audience.
- Record brands, visible sources, dates, missing answers, and failed checks.
- Choose one information, access, or evidence problem to investigate. Record the change before repeating the checks.