AEO vs GEO vs AI-Mediated Shortlisting
AEO and GEO track citations, mentions, and recommendations across time.
AI-mediated shortlisting matches specific buyer requirements against vendor capabilities.
Every test is a simulation of a private buyer conversation, and how the simulation is built decides how accurate the answer is.
AEO, GEO, and AI-mediated shortlisting are three ways to test what AI says about a company when a buyer asks. The tests look for one of four outcomes:
A citation: AI links to the company's website as a source, like a footnote
A mention: AI lists the name of a company in its answer
A recommendation: AI includes the company in a set of companies to consider
The shortlist: AI names the vendor as a match for all of the buyer's stated requirements, and says why
"Recommendation" and "shortlist" are often thought of as the same thing. The difference is the level of detail.
A recommendation is general: AI names a company that fits a market, a type of buyer, or a topic.
A shortlist is specific: AI names a vendor that fits all of one buyer's stated requirements, and says why.
How Do AEO, GEO, and AI-Mediated Shortlisting Differ?
AEO, GEO, and AI-mediated shortlisting test what AI says about a company, but each one is built to improve a different outcome.
AEO and GEO are built to improve visibility.
AI-mediated shortlisting is built to increase the probability that AI returns an accurate answer.
AEO | GEO | AI-mediated shortlisting | |
|---|---|---|---|
Goal | Visibility: be cited or mentioned in the answer more often | Visibility: be mentioned, cited, or recommended more often | Accuracy: make the shortlist, with the reason why, when a buyer tells AI their requirements and asks AI to match them against vendor capabilities |
Where | Google AI Overviews, and AI assistants such as ChatGPT, Claude, Gemini, Copilot, and Perplexity | Google AI Overviews, and AI assistants such as ChatGPT, Claude, Gemini, Copilot, and Perplexity | AI assistants such as ChatGPT, Claude, Gemini, and Copilot |
What it measures | How often the company is cited or mentioned in the answers to topic prompts about a category, tracked over time | How often the company is mentioned, cited, or recommended in the answers to topic prompts about a category, tracked over time | If the vendor makes the shortlist for a buyer who states their requirements. Why AI says the vendor is a good fit |
What Are the Prompts Based On?
Because the conversations buyers have with AI are private and cannot be observed, prompts must be simulated. There are seven data sources a simulated prompt can be built from. Choosing a source is a trade-off between who the source describes and the effort it takes to produce the prompts.
Prompt source trade-offs
Source | Who the source describes | Effort to produce the prompts |
|---|---|---|
1. The company's website | A typical buyer | Low |
2. A category or topic | A typical buyer | Low |
3. Search keywords, rephrased as questions | A typical buyer | Low |
4. Panels of real AI users | A typical buyer, in real people's words | Low |
5. The company's own records Sales call transcripts, support tickets, CRM notes | The company's buyers | Medium |
6. Interviews with the company's CEO or sales team About the company's best clients | The company's buyers | Medium to high |
7. Interviews with the company's customers About the customer's own purchase | The company's buyers, in the buyer's own words | High |
Sources that describe a typical buyer
1. The company's website
A tool generates the prompts from what the company says about itself, which marketing wrote, and a person reviews the prompts.
2. A category or topic
The prompts describe a typical buyer in the category, not the company's buyer. A tool suggests the prompts and a person reviews them.
3. Search keywords, rephrased as questions
The keywords are real search demand, but the keywords start from what people type into a search engine. The keywords have the topic, not the buyer's situation. A tool converts the keywords into prompts and a person reviews the prompts.
4. Panels of real AI users
The prompts are real people's words, but the people are anonymous and not targeted on the company's buyers. The tool provider supplies the data.
Sources that describe the company's buyers
5. The company's own records
Sales call transcripts, support tickets, and CRM notes. The buyers' own words, but recorded at the wrong point in the purchase. A sales call is usually recorded after the shortlist is built, in Job 3: Validate Vendors, when buyers ask validation questions, not shortlisting questions. Support tickets and CRM notes are recorded after the purchase, when the buyer is a customer with a problem, not a buyer with requirements. The records have to exist, be sorted, and be prioritized, because the records can include buyers who are not a fit along with the best ones.
6. Interviews with the company's CEO or sales team
An interview about the company's best clients. The company's buyers, in the team's account of them. The interview takes a few hours of the team's time.
7. Interviews with the company's customers
An interview with each customer about the customer's own purchase. The buyer's own account, though recalled after the fact. Each customer has to agree to an interview and schedule it.
What Does a Prompt Contain?
A prompt is a one-turn conversation with AI: a buyer asks a question, and AI answers. A prompt can contain a simple topical question, a buyer persona with a question, or the buyer's requirements with a request for a shortlist.
Topic prompt
A simple topical question about a category, usually under ten words, that might include some qualifiers or even some high-level requirements. For example:
Best CRM software
Best CRM software for manufacturing
Best CRM software for a 60-person metal fabrication shop that costs under $15K a year
Qualifiers and high-level requirements narrow the group of buyers: an industry, a company size, a budget.
Strength: quick and cheap to produce, so many can be tracked across a group of buyers over time.
Limit: a topic prompt is a question about a group of buyers, not a question from a specific buyer. Qualifiers and high-level requirements make the group smaller, but do not include the specific requirements from one buyer.
Persona prompt
A prompt that gives AI some context about the type of buyer before asking a question. The context is a buyer persona: the vendor's description of a typical client, with the client's role, kind of company, goals, and problems. For example:
Persona: sales manager at a mid-sized metal fabrication shop, 50 to 100 employees, with three to five salespeople. Goals: win more repeat work. Problems: custom quotes are tracked in spreadsheets and follow-ups get missed.
Question: What is the best CRM software for a shop like this?
Strength: everything a topic prompt gives AI, plus the goals and problems of a type of buyer. A persona is written once and can be used to ask several different questions, so persona prompts are still relatively quick to produce and simple to track over time.
Limit: a buyer persona is still a group of buyers, smaller than the group a topic prompt describes, but not a specific buyer. A persona is written by the vendor, so the goals and problems in the persona are what the vendor believes a typical client has, not what a buyer said.
Requirements prompt
A prompt in which a specific buyer describes the buyer's own situation and asks AI for a shortlist. The situation is the buyer's requirements: who the buyer is, what problem the buyer is trying to solve, and what the buyer needs from a vendor. For example:
ok I need help picking a CRM. we're a metal fab shop in Ohio, about 60 people, 4 guys in sales. we quote 40-50 custom jobs a month and it's all living in spreadsheets, so stuff falls through the cracks constantly. last week we lost a repeat customer because nobody followed up on a quote. I'm done with it. needs to track a job from quote to order, work on the shop floor and on a phone when the guys are out at customer sites, and I'm not paying more than 15k a year. what should I be looking at and why?
Strength: everything a persona prompt gives AI, plus the requirements of one buyer in the buyer's own words. AI has a specific buyer to compare vendors against.
Limit: a requirements prompt is one buyer's wording, and another buyer with the same needs would say it differently. Writing one needs a source that knows the buyer, so requirements prompts take more effort to produce and are harder to track over time.
How Many Turns Does a Test Take?
A simulated buyer's conversation can be one prompt or several. A buyer's real conversation with AI usually takes more than one turn, because a buyer rarely states every requirement in the first turn. The buyer reads the answer, then adds a requirement, rules a vendor out, or asks a follow-up question.
Single turn
One prompt and one answer.
Strength: a single-turn prompt is simple to repeat, so results can be compared over time.
Limit: the simulation begins with one prompt and ends with a single answer; there is no additional follow-up. One question and one answer is also not how buyers use AI in their purchasing process. Forrester found that buyers expect AI to understand their intent and guide them through complex decisions, and that buyers use AI to evaluate the differences between vendors and to justify the purchase. Guiding a buyer through a complex decision takes more than one turn.
Multi-turn
A scripted conversation: several prompts given to AI in sequence, with the follow-up prompts written by whoever runs the test, either in advance or by an AI agent during the test.
Strength: a multi-turn test is a closer simulation of how buyers use AI. A buyer adds a requirement, rules a vendor out, or asks a follow-up question after reading each answer, and a multi-turn test sends those turns.
Limit:
The follow-up prompts come from whoever runs the test, not from a buyer, so every follow-up is a guess at what a buyer would say next.
Non-determinism has a bigger effect across turns. AI's answer to the first prompt varies from run to run, so every later prompt lands on a different answer, and the chance that AI returns the same final answer decreases as the number of turns increases. A study of 15 AI models found that the same information given across several turns instead of in one prompt produced less reliable answers.
The complexity of testing every possibility increases with the number of turns. If each turn adds one requirement, two turns can play out in 2 orders, three turns in 6, four turns in 24.
Comparing AI models is harder. In a single turn, ChatGPT, Claude, Gemini, Copilot, and Perplexity all receive the same prompt, so the answers can be compared. In a multi-turn test, each model gives a different first answer, so the second prompt lands on a different answer in each model, and from that turn on, each model is in a different conversation.
About the author
David Lee
David Lee is the founder of Do What Works. At Toyota Motor Sales, he was part of a team that selected a vendor, from gathering requirements to the final decision. He then ran a B2B marketing agency for ten years. Today he runs buyer conversations through ChatGPT and Claude to see how AI researches and shortlists vendors.