Short answer: to get recommended by ChatGPT, three things have to hold at once. ChatGPT's search agents can reach your pages, your pages state comparable facts about you that can be lifted into an answer, and the same facts appear somewhere independent of you. Being cited as a source is not the same outcome as being recommended, and the work that produces one does not automatically produce the other.

No one outside OpenAI controls which brands an answer names, so nothing below is a guarantee. It is the list of preconditions you can control, five steps in the order they usually block, and a way to measure whether they are working without fooling yourself.

The vocabulary matters, because each outcome has a different cause and a different fix:

OutcomeWhat it meansWhat moves it
MentionedYour name appears somewhere in the answerA clear, consistent entity; presence on other sites
CitedOne of your URLs is marked as a sourceCrawler access, server-rendered text, liftable passages
RecommendedThe answer actively suggests you for the buyer's situationComparable claims, corroborated by independent sources

A brand can be cited without being recommended: ChatGPT uses your page as reference material and then suggests a competitor. It can also be recommended with no link to it at all, because the answer leaned on review sites and round-ups instead of your pages. If your reporting merges the three into one number, it hides exactly the gap you need to see.

STEPOUTCOME IT MAINLY MOVES1 Let the right agents inOAI-SearchBot and ChatGPT-User reach you2 Answer a specific situationA constraint, a price band, a use case3 State comparable factsLiftable sentences, limits included4 Get corroboratedThe same facts on pages you do not own5 Be one entityOne name, one description, sameAs linksMentionedYour name appears in the answerCitedYour URL is a marked sourceRecommendedThe answer suggests you
Figure 1. The five steps in this article and the outcome each one mainly moves. Access and liftable passages lead to citations; situations, facts and corroboration lead to recommendations. Source: InsightWonder measurement rules and mechanics described in this article.

The five steps below map onto those outcomes unevenly. Access and liftable passages mostly decide whether you are cited. The situation you target, the facts you state and the corroboration you earn decide whether you are recommended. A consistent entity decides whether the mentions you do get are counted as one brand.

Step 1: let the right ChatGPT agents in

OpenAI documents three agents, and they do different jobs. GPTBot crawls content that may be used to train OpenAI's models. OAI-SearchBot builds the index that surfaces websites in ChatGPT's search features. ChatGPT-User fetches a page because a person in ChatGPT asked for it. Blocking GPTBot is a defensible position on training and does not remove you from ChatGPT search results. Blocking OAI-SearchBot does, and it is the agent sites most often block by accident while meaning to opt out of training.

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: ChatGPT-User
Allow: /

Then test it with the real user agent rather than trusting the file: a CDN or firewall rule that returns 403 to these agents is a block in practice, whatever robots.txt says. Check that the page text is in the HTML the server sends, too, because most AI crawlers do not execute JavaScript. The primary source is OpenAI's crawler documentation, and OpenAI also publishes an IP list for each agent, such as searchbot.json for OAI-SearchBot, which is what a firewall allow rule should use. Our walkthrough of blocking GPTBot without disappearing covers the edge cases.

Step 2: be the answer to a specific situation

Recommendations are answers to situations, not to category names. A question that carries a budget, a constraint or a use case produces a different set of named brands from the bare category term. "Best project management software" and "project management software for a five-person agency that bills hourly" are different contests.

This matters most where a category already has a default answer. In our own Q3 2026 benchmark, the three categories with a clear default were all product categories: the leading brand appeared in 38% of project management software answers, 35% of skincare answers and 28% of electric bike answers. The decision that follows is the same in any category:

From category term to a situation you can winCATEGORY TERMproject managementsoftware+ WHOfor a five-personagency+ CONSTRAINTthat bills itsclients hourlyA different contest,different brands namedDoes one brand already appear in a quarteror more of the answers to the category term?YESNOCompete on a sub-question theincumbent answers badlyOur benchmark: all three such categories were productsThe head term itself isstill open groundOur benchmark: every local-service category
Figure 2. Adding who is asking and what constrains them turns a category term into a different contest; whether to compete on the head term depends on how consolidated it already is. Source: InsightWonder Q3 2026 AI recommendation benchmark, 12 categories across five models; the example query is from this guide.

When one brand already appears in a quarter or more of answers, competing on the head term is expensive. The more realistic route is to become the answer to a sub-question the incumbent answers badly. Every local-service category in the same benchmark had a leader below 13%, so in those categories the head term itself is still open.

Choosing the situations to target

Write down the situations your best customers were in when they chose you, then turn each into a question a stranger would ask. Useful dimensions are the ones an answer has to resolve before it can recommend anything:

  • Constraint: team size, compliance requirement, integration, dimensions, compatibility.
  • Price band: a budget the buyer states, or a billing model such as hourly or per seat.
  • Use case: the job the product is bought for, described the way the buyer would describe it.

Questions where ChatGPT currently names nobody are open ground and cheaper still than displacing an incumbent. They are worth finding first.

Steps 3 to 5: give ChatGPT grounds to recommend you

Step 3: state comparable facts, including where you lose

"Best X for Y" forces the engine to line candidates up on shared dimensions: price, coverage, who it is for, what it integrates with. If your pages state none of those in sentences that can be lifted, you are background reading rather than a candidate.

  • Facts, not adjectives. Assistants lift sentences that answer a constraint: a dimension, a material, a return window, a compatibility. Copy written for brand tone gives them nothing to lift.
  • Sentences, not only tables. Put the key specifications into prose on the page as well as into a spec table, so a single passage can carry them.
  • Say who you are not for. A page that states a limitation plainly is easier to place in a comparison than one that claims to be best at everything.

The habits that make a passage liftable are covered in writing for extraction.

Step 4: get corroborated where you do not control the page

A model needs to see the same fact from independent places before it will state it confidently. A claim that exists only on your own domain is the weakest evidence you can offer, however well it is written or marked up. That is why corroboration is usually the bottleneck between being cited and being recommended.

The work is unglamorous. In product categories a large share of citations go to marketplaces, review sites and publisher round-ups rather than to brand sites, so check that those pages carry your current facts, prices and availability. For brands, the common failure is not silence but a plausible wrong detail, such as an old price or a market you left, which usually traces back to one stale external source. Correct those sources first. Third-party corroboration explains how to find and prioritise them.

Step 5: make the brand one entity

Spelling variants, a product line described as a separate company, and a legal name that differs from the trading name can each turn one brand into several in an answer, and in your own measurement. Write one plain description of what you are and who it is for, use it everywhere you control, and publish Organization markup with sameAs links to the profiles that are really yours. Entity and sameAs covers which links to include.

Some popular tactics either have no evidence behind them or actively damage the result. The one controlled study we can point to is the GEO paper by Aggarwal and colleagues, accepted to KDD 2024, which tested content changes against generative engine responses in a research setting rather than against ChatGPT itself:

TacticWhat the evidence saysVerdict
Keyword stuffingLittle to no improvement in the GEO paperSkip it
Citing sources, adding quotations and statisticsThe best-performing methods in the GEO paper, with visibility gains of up to 40% in its testsDo it, with real, checkable sources
Seeded praise and planted reviewsDoes not create independent corroboration, and is easy to spotAvoid
One rule blocking every OpenAI agentRemoves you from ChatGPT search resultsBlock by name and role instead
Measuring with brand-named questionsInflates the score without any change in answersExclude those questions

The pattern is that anything which only makes a page look more relevant does little, while anything that makes a claim easier to verify helps. That matches the rest of this article: recommendations follow evidence that can be checked from more than one place.

Measure mentions and recommendations separately

Track mention rate and recommendation rate as separate columns over a fixed set of questions, per engine, before any average. A brand that is mentioned often but rarely recommended has a corroboration problem. A brand recommended without being cited has an attribution gap. One number cannot tell you which you have.

Two measurement errors flatter brands more than any other. Questions that contain your brand name get answered with your brand name, so exclude them. And competitors must be counted the same way you are counted. In one of our own incidents, asymmetric detection showed one brand at 31% share of voice and second place; counted symmetrically, the same answers put it at 0% and last. The whole difference is in how the names are counted:

One answer, two ways of counting the brands in itAnswer: “For a small team, try Northwind or Contoso. Acme Inc. also works.”ASYMMETRICWhat inflated our own reportYou: matched looselyAcme, Acme Inc., acme.com all countCompetitors: a fixed list onlyNorthwind is on it; Contoso is not, so it is missedYour share rises by construction,whatever the answer actually says.SYMMETRICHow every brand is counted nowOne matcher for every brandthe same variants rule for you and for themNames come from the answerAcme, Northwind and Contoso are all foundYour share is what the answers say:the only version worth tracking.
Figure 3. One answer counted two ways. The brand names are placeholders; the asymmetric version is how our own share-of-voice report counted until we fixed it. Source: InsightWonder share-of-voice incident, fixed 2026-07-20.

When a competitor is named and you are not, record that it happened, not a guess at why; the data does not tell you why ChatGPT chose it. Only trust a change that repeats in two consecutive runs, because answers vary even when nothing has changed. Building a prompt set explains how to write the question set, and our ChatGPT engine page summarises what we have measured there.

Sources