Chapter two of ten
An engine answers from memory or through retrieval, a live search of the web. Retrieval is the part you can shape.
Buyers ask Does ChatGPT already know us, or does it look us up?
By Maeva Cifuentes · Updated October 2026
What the model learned in training
What it found a second ago
The short answer
An answer engine has a memory and a browser. Models build a memory from their training data. The labs refresh that data every few months, which puts it outside our real-time influence. Retrieval is the live search the engine runs before it answers. It drives buying questions and it responds to work you can do this quarter, so this guide focuses on retrieval.
Maybe you have seen this before. An engine quotes your page as a source and then recommends three competitors. Or it names you on Monday and leaves you out on Tuesday. Both come from how the engine builds an answer.
Picture a well-read expert with a browser. For some questions the expert answers from what they already know. That is training data, which the model absorbed months ago and which stays frozen until the next model ships.
For other questions the expert opens the browser, reads a handful of pages and then answers. That is retrieval, the term this guide uses for an engine's live search. Google's own developer documentation says the model decides for itself whether a search would improve the answer.
How often the engine reaches for sources
100 buyer-journey prompts run twice through GPT-5.2. Source Growth Memo, Reasoning lift, using Semrush data. A separate academic study of logged conversations found GPT-4o answering about 24% of prompts with zero site visits, against about 2% for Perplexity's model (Strauss et al.).
Try it
Pick a question a buyer might ask and see how an engine tends to handle it.
Typical behavior drawn from the studies cited in this chapter. Engines and settings vary.
Buying questions tend to trigger retrieval, so the money you spend on being findable and well described pays back where buyers choose.
Ask your buyers' questions with web search turned off, then on. The difference shows what the model remembers and what it finds.
A brand can drop out at any of retrieval's four steps.
Fresh, specific and commercial questions push the engine toward the browser.
One prompt becomes several smaller searches. Google calls this query fan-out.
Each search returns a set of pages from the engine's index.
The engine retrieves far more pages than it cites, then writes one answer.
In The Search Spot study of 488 product pages, ChatGPT ran a category search first and then searched for vendors by name. Pages ranking in Google's top 10 for five or more searches were cited regularly 76% of the time, against 10% for pages below that mark. Brands that rank in the top 10 for a wider set of relevant keywords are far more likely to be cited.
Chapter 3 goes deep on the splitting step and how to see the searches an engine runs for your category.
more searches with high reasoning than with minimal, 1,130 against 245 across 100 prompts
Growth Memo with Semrushmore pages retrieved by ChatGPT than cited. 85% of retrieved pages go uncited.
AirOps, 548,534 retrieved pagesmedian pages Perplexity's model visited and then cited per answer
Strauss et al., 2025Step four, picking
Share of retrieved pages that ChatGPT cited, by position in its search results. Source Growth Memo with AirOps, 815,484 query and page pairs. Bar charts in this chapter share one scale from 0 to 100%.
Being cited and being mentioned are separate wins. Ask your team which step the brand is dropping out at before funding more content.
Check each step in order. Confirm the bots can reach the page, confirm it ranks for the smaller searches, then look at what the page says.
Important definitions for you to understand
One answer, two outcomes
For a 500-person healthcare company, the strongest options are Rival A and Rival B.
Here your page is cited and your brand goes unmentioned. The reverse also happens, where an engine names you from memory and links to someone else's page.
We treat mentions as the main goal and citations as the path to them. Chapter 8, on measurement, shows how to track each one.
Ask for both numbers in every report. A high citation count with few mentions means the engines use your content to recommend someone else.
Log cited pages and mentioned brands as separate columns for each prompt, so the two trends stay readable.
Both memory and retrieval shape an answer. We focus on retrieval, because it responds to work within weeks and it carries the buying questions.
The two lists connect. What the web says about you today is what a future model learns from, so retrieval work done now becomes memory later.
Rankings carry a lot of the load. AirOps found that 43.2% of pages ranking first on Google were cited by ChatGPT, 3.5 times the rate of pages beyond the top 20.
Fund the left-hand list and judge the team on it. Treat the right-hand list as weather, worth watching and pointless to fight.
Turn the left-hand list into your audit. Each line maps to a later chapter, from technical access to off-site mentions.
The engines share the same four steps and differ in where they search and how much they show. Treat each one as its own channel.
| ChatGPT | Google AI Overviews and AI Mode | Gemini app | Perplexity | |
|---|---|---|---|---|
| Where it retrieves from | Third-party search providers plus content from partners. April 2025 court testimony named Bing and described OpenAI's own index as still in development. | Google's search index, with several related searches per question. | Google Search, when the model judges a search useful. | Its own search index. |
| Bots that feed retrieval | OAI-SearchBot builds the index. ChatGPT-User fetches a page live. | Googlebot. | Googlebot. | PerplexityBot builds the index. Perplexity-User fetches a page live. |
| Bot that feeds training | GPTBot. | Controlled through Google-Extended. | Controlled through Google-Extended. | Perplexity says its bots stay out of model training. |
| What a page needs | OAI-SearchBot allowed in robots.txt. Blocked sites drop out of search answers. | Indexed and eligible for a snippet. Google asks for zero special files or markup. | The same as Google Search, with Google-Extended allowed, since that token also controls grounding in Gemini Apps. | PerplexityBot allowed in robots.txt. |
| How it shows sources | Cited a source in about three of four logged answers. | Links sit beside the answer and count in Search Console under Web. | Left 92% of logged answers without a citation. | Visited a median of 10 pages and cited 5. |
| What to watch | Hard questions get routed to a thinking mode that searches far more. | AI Mode works as a closed loop where buyers accept the shortlist. | Personal history shapes the answer. | It searches on nearly every question. |
Bot roles come from OpenAI, Google and Perplexity documentation. Citation figures come from Strauss et al., who studied about 14,000 logged conversations with GPT-4o, Gemini and Perplexity's Sonar in March and April 2025. Consumer apps may behave differently. The authors also note that OpenAI's logs look trimmed.
Ask which engines your buyers use before spreading effort across all of them. Chapter 1 showed that 91% of cited pages appear in one engine only.
Open robots.txt and check the search bots by name. A training block on GPTBot leaves search untouched, while a block on OAI-SearchBot removes you from answers.
Ask the same question twice and you can get two different shortlists. Rand Fishkin of SparkToro and Patrick O'Donnell of Gumshoe had about 600 volunteers run 12 prompts close to 3,000 times. The chance of ChatGPT or Google returning the same list of brands twice came in under 1 in 100, and the same list in the same order came in near 1 in 1,000.
A trend still shows through the variation. That trend is what we measure. Three things cause the variation.
A model picks its words with a little randomness built in, so each answer is one draw from many possible answers.
The engine may split the question differently each time and so read different pages.
Reasoning mode, location and personal history all shift the result.
One answer is a sample. Track a category or theme of prompts and read the direction of your share across many runs.
The trend under the noise
Volunteers each wrote their own prompt for buying headphones, and almost none resembled another. The 994 answers still named Bose, Sony, Sennheiser and Apple in 55% to 77% of cases. A broader B2B category showed lower shares, at 30% to 40% for the top brands. Source SparkToro with Gumshoe, January 2026. The author flags a small sample and a short time window.
That is why we track prompts at the level of a category or theme. Buyers phrase things in endless ways, so a single prompt tells you little, while a group of prompts on one theme shows which way you are moving.
Source Growth Memo with AirOps, 16,851 queries run three times each through the ChatGPT interface, with each page scored on how often it was cited across every query where it appeared. Segments share one scale from 0 to 100%.
Reasoning mode moves the result further. In the Growth Memo test, only 25.6% of cited domains appeared under both minimal and high reasoning.
A screenshot of one answer proves little, whether it flatters you or alarms you. Ask for the trend by category or theme across a fixed set of prompts.
Group prompts by theme and report the direction of each group. Chapter 8 covers how tracking tools turn repeated runs into a share.
Ethan Smith of Graphite says AEO works on retrieval, since training data is hard to influence. We agree on where to focus.
Kevin Indig argues that what a model learned about a brand in training colors which retrieved pages it chooses to cite.
Eli Schwartz expects personalization to spread from Gemini to AI Mode, then to AI Overviews and the rest of Google Search, so two people asking the same thing get different answers. We read that as one more reason to track themes over single prompts.
Google's documentation says pages need zero extra optimization to appear in AI Overviews or AI Mode.
Choose your next stop
We use analytics cookies to see which pages help people. They load only if you accept. Cookie policy