How to Get Recommended by ChatGPT: The Five-Rung Ladder, and How to Tell Which Rung You Are Stuck On

Getting recommended by ChatGPT is a sequence, not a switch. The model has to be able to fetch your site, lift a self-contained answer from it, resolve your name to a single business, find that claim corroborated somewhere it already trusts, and see it hold across repeated prompts. Skip a rung and the ones above it cannot carry weight.

A homeowner asks ChatGPT which roofer to call in their city. A founder asks it to shortlist three outbound agencies. Neither sees ten blue links. They see two or three names, and they act on them. There is no advertising slot in that answer, no submission form, and nobody at OpenAI to email about it.

That surface is not experimental any more. Google made AI Overviews generally available in 2024, ChatGPT Search shipped in October 2024, and Perplexity has cited its sources by default since launch. What follows is the sequence that decides whether your business appears in those answers, in the order the sequence actually runs.

Why does ChatGPT recommend competitors instead of you?

Almost always for mechanical reasons rather than editorial ones. The model could not fetch your pages, could not find a self-contained claim to lift, or could not confirm that the name on your site and the name in its sources refer to the same company. Preference comes last, not first.

The ladder, in order

Each rung depends on the one below it. Fixing rung four while rung one is broken changes nothing, which is why so much AI visibility work produces no movement.

1. Access. The relevant crawlers can fetch a rendered page and read real text.
2. Extractability. A specific claim sits somewhere a machine can lift it cleanly.
3. Identity. Your name resolves to one business, not three near-matches.
4. Corroboration. Sources the model already trusts repeat the same claim.
5. Recurrence. You hold across repeated runs and differently-worded prompts.

Rungs one and two are edits to pages you already own, and they are cheap. Rungs three and four accrue over months. Rung five is measurement, and it is the only one that tells you whether the other four worked.

Can ChatGPT actually reach your website?

Check before assuming. OpenAI publishes three separate user agents: GPTBot for training, OAI-SearchBot for search indexing, and ChatGPT-User for live retrieval. A robots.txt rule written in 2023 to block training will usually block the other two as well, and many sites are excluded by a line nobody remembers adding.

The crawlers that decide whether you exist

Nine user agents carry most of the weight in 2026. Blocking any one of them removes you from that system’s view of the web, and the rules are documented publicly by OpenAI and Google:

• GPTBot, OAI-SearchBot, ChatGPT-User — OpenAI, three distinct jobs
• ClaudeBot — Anthropic
• PerplexityBot — Perplexity
• Google-Extended — Gemini and AI Overviews, introduced in September 2023
• Applebot-Extended — Apple Intelligence
• CCBot — Common Crawl, an input to many models
• Bytespider — ByteDance

The second failure at this rung is quieter. A client-side rendered site returns HTTP 200 with an empty container: nothing errors, and the crawler simply reads a page with no headings, no prose and no facts. Fetch your own page without JavaScript and read what comes back.

What makes an answer liftable?

Self-containment and length. The passage has to make sense quoted with nothing around it, sit immediately under a heading phrased the way a buyer asks, and run roughly forty to sixty words. Shorter loses the context a model needs; longer gets truncated in favour of a competitor who was tighter.

The forty-to-sixty word rule, applied

This is the standard DaxReach holds its own pages to, and the reasoning is set out in more detail in what answer engine optimization actually is:

1. A question-shaped heading that ends in a question mark or opens with who, what, why, how, is, does, should or which.
2. A 40 to 60 word direct answer immediately beneath it, before any sub-heading interrupts.
3. A sub-heading to close the block, after which you can write at whatever length the subject deserves.
4. Four facts per hundred words across the page: figures, dates, named sources and outbound links a model can follow.

Point three is the one most writers miss. A question followed by 160 words of excellent prose is harder to quote than one followed by 47 words and a sub-heading, because the unit a model would lift is too long to use whole.

How does a model know your business is one entity?

By corroboration across independent sources. If your site, your LinkedIn page, your Google Business Profile and any directory listing all state the same name, category and location, the model resolves one entity. When they disagree, it either picks the most-cited version or declines to name you at all.

Entity signals a model can actually check

None of these are exotic, and all of them are verifiable from outside your company:

• A consistent legal and trading name across every profile you control
• The same category wording on your site and your third-party listings
• A sameAs list in your Organization markup pointing at profiles the page genuinely links to
• Matching contact details, because a mismatch reads as two different businesses
• Independent pages that describe you the same way you describe yourself

A sameAs entry pointing at a profile your site does not link is worse than none: it asserts a relationship the crawler cannot confirm. This rung is where most young domains lose, and it cannot be bought or rushed.

Which rung are you stuck on?

Diagnose by symptom rather than by guesswork. Each failure mode looks different from the outside: absent from every answer, named but described wrongly, mentioned only when you supply the brand name, or recommended in one run and forgotten in the next. Match the symptom to the rung below.

The diagnostic table

The last row is the one that misleads people most often. Assistant responses vary between runs, so a single flattering answer is not evidence of anything. Run each prompt several times before drawing a conclusion.

How do you measure whether any of this worked?

Run the same buyer prompts again and compare. Score the three surfaces separately: classic search ranking, extracted answers, and generative recommendations. Averaging them hides the failure, because a business can rank on page one and still be absent from every assistant answer that matters.

The signals the GAS checker reads

DaxReach’s checker grades a domain on readiness signals it can verify without guessing, and reports them as three separate scores rather than one blended number. It reads crawler permissions across those nine user agents, whether an llms.txt file exists, whether your question headings are answered inside the extractable range, whether Organization and FAQ structured data is complete, and whether your profile signals agree with each other. The full scoring method is set out in the GAS Engine playbook, and the vocabulary in the glossary.

Readiness is not the same as being recommended. It measures whether the machine can use you, which is the part you control. The recommendation itself depends on rungs four and five, which take time.

Three routes to getting named, compared

An honest note on that last row: a good traditional SEO agency will beat a structure-first approach on competitive commercial head terms, because links still decide those. The argument is not that ranking stopped mattering. It is that ranking stopped being sufficient, and the two jobs are no longer the same job. If the constraint is strategy rather than execution, a fractional CMO is the better-fitting engagement, and agencies reselling this work have their own partnership route. Figures for all of them sit on the pricing page.

The businesses being named by assistants in 2026 are not publishing more than everyone else. They fixed the bottom two rungs, stopped guessing about the top three, and measured the same prompts twice.