I do a lot of discovery calls, and three questions come up on a lot of them. What can I do to rank higher for AI? How do I rank on ChatGPT? What’s your system for finding where I’m ranked on ChatGPT?

The short answer I give everyone is the same one. You can’t rank in ChatGPT the way you rank in Google. There’s no positioning system underneath it, nothing like Search Console or Bing Webmaster Tools, no position three to hold, no list to climb, nothing to watch move week to week.

That’s not the whole answer, though. ChatGPT does rank what it retrieves. It just never shows you the order and never promises you a place in it. OpenAI says so itself, in plainer words than most agencies will use with you: “Ranking in ChatGPT Search is based on a number of factors designed to help users find reliable, relevant information. There is no way to guarantee top placement.”

So there is ranking, and there is no rank you can track. Only one of those is a problem you can do anything about. Here’s what you can control, and how to check it yourself in half an hour.

01Ranking and Retrieval Are Not the Same Thing

Google shows you an ordered list. It shifts with location, device and personalisation, but there’s a position there, it’s roughly stable, and you can watch it. ChatGPT assembles an answer per conversation out of whatever it retrieved, some of it read on the spot and some of it pulled from copies it already held. Two clients in the same town can ask the same question and get different practices named.

So “how do I rank in ChatGPT” is the wrong question with the right instinct behind it. The questions that have answers are: can ChatGPT fetch my page, and does that page say something worth quoting?

02The Three Routes That Carry Web Content to You

Three routes carry web content to a practice. A fourth surface runs on business listings rather than web pages, and section 04 deals with that separately. Treat this as a working model, not a full map.

Training data. The model absorbed a snapshot of the web during training. Whether your site made it in, or survives in a recoverable form, isn’t something you can find out. You can’t submit to it, correct it, or hurry it. Treat this path as weather.

Live web search. When ChatGPT searches, it draws on an index of pages it has already collected, alongside outside search providers. Getting into that index is what you can work on, and OpenAI is explicit about the veto: “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.” Note what that does and doesn’t say. Opting out reliably keeps you out. Opting in makes you eligible, and OpenAI publishes no formula for what happens next.

A direct fetch. When someone pastes your URL into a chat, an agent called ChatGPT-User requests it, though what comes back may be a stored copy rather than a fresh read. OpenAI’s documentation notes something worth knowing: “Because these actions are initiated by a user, robots.txt rules may not apply.” Blocking the crawlers doesn’t reliably stop this one.

Two of the three are workable, and the middle one is where the work lives.

03Does ChatGPT Use Google Results, or Bing?

Partly Google, in some modes. Not Bing, in the way people assume.

For years the shorthand was that ChatGPT sat on top of Bing. In July 2026 the French consultancy Resoneo tested that by reverse-engineering what ChatGPT actually fetched across 1,249 answers, roughly 88,000 search results and 26,900 distinct pages. Writing it up in Search Engine Land on August 17, 2026, Olivier de Segonzac reported four retrieval pipelines, one of them an OpenAI in-house index labelled labrador. That’s observed behaviour rather than architecture OpenAI has confirmed, and it’s one month’s snapshot.

The Bing finding is blunt. “Only 1.5% of labrador URLs appear in Bing’s top 20 for the same fan-outs.” His conclusion: “OpenAI has its own web index, and Bing doesn’t supply it.” That’s narrower than “ChatGPT doesn’t use Bing,” since OpenAI still names Microsoft as a search partner in its own help docs. What the study rules out is Bing being the thing underneath.

Google is a different story. Reporting the same research in Search Engine Journal on August 11, 2026, Matt G. Southern noted that in paid thinking mode roughly 75% of results came from scraped Google results, against about 24% from the in-house index. On free accounts the in-house index handled most.

Mostly good news for a small practice, then. Your ordinary Google SEO still feeds the paid mode, and the free mode leans on an index that isn’t limited to OpenAI’s licensed publishing partners.

One caveat: both write-ups cover the same Resoneo dataset. That’s one study reported twice, not two confirmations.

For the wider picture across Google, Copilot and Perplexity, the GEO vs SEO pillar covers what each system does differently.

04"Therapist in [City]" May Never Touch Web Search

Here’s the part that complicates everything above, and the reason this post can’t stop at crawler settings.

The same Resoneo study found that some question types skip web search entirely. In Search Engine Land’s words: “Shopping and local never touch web search. If you sell products or run a restaurant, your battle is fought in merchant feeds and business listings, on top of the classic citations.” The feeds named behind local answers are Yelp, TripAdvisor and Google Maps.

The study’s prompt themes did include health, but it doesn’t report a therapy-practice-specific local sample, so I can’t tell you the exact route a clinical query takes. What I can say is that “find me a therapist in Chicago” is the same shape of question, and if it routes the same way, your website isn’t the asset being read. Your business listing is.

So treat this as a parallel bet, not a replacement for the rest of the post:

• Claim and fill out your Google Business Profile properly. Hours, categories, service area, the lot.
• Keep your directory profiles current and consistent with your site. For most clinicians I’d check Psychology Today first.
• Make sure your name, address and phone number are accurate and consistently represented everywhere they appear.

That’s ordinary local SEO for therapists work and it was worth doing before ChatGPT existed. The new part is that it may be feeding a pipeline your website can’t reach.

05The Five Things That Decide Whether ChatGPT Can See You

Let the crawler in. Your robots.txt needs to allow OAI-SearchBot. You can allow it while still disallowing GPTBot, which is the training crawler, if you’d rather not feed model training. OpenAI treats them as separate decisions. Its documentation notes it can take around 24 hours for a robots.txt change to register.

Let the crawler’s IP addresses in. This one fails silently, which is what makes it worth checking by hand. OpenAI’s guidance says to “ensure your site host and/or content delivery network allows traffic from our published IP addresses.” Those ranges are published as JSON files at openai.com/searchbot.json, openai.com/gptbot.json and openai.com/chatgpt-user.json. A permissive robots.txt means nothing if Cloudflare, a security plugin or your host’s firewall is turning the crawler away at the door.

Sort out your location signals. ChatGPT estimates general location from IP address and may rewrite “near me” into a city query. OpenAI’s own example turns “good restaurants near me” into “top restaurants San Francisco.” So the query reaching the index is likely to be “therapist in [city]” rather than “therapist near me”. Your city needs to be on the page in plain text, not just implied by an address block.

Link to your new pages from somewhere. Resoneo found OpenAI’s discovery running on links rather than sitemaps. GPTBot pulled their sitemap daily and followed none of the URLs listed only there, while crawling every page reachable by link. A page nothing links to may never be found.

Publish pages that answer the question being asked. Not brochure copy about your approach. The literal questions: what you treat, who you work with, whether you take insurance, whether you have openings. How to get cited by AI search goes deeper on the structure side.

06The Exact Prompts to Test Yourself With

Open ChatGPT logged out, in a normal browser window, and run these as written, substituting your city and speciality. Logged out matters, because OpenAI says saved memories can be used to rewrite your query, and your own account knows who you are.

• Who are the best therapists in [city] for [speciality]?
• I need a therapist in [city] who takes [insurance]. Who should I call?
• Is [your practice name] a good therapy practice?
• Does [your practice name] have openings right now?
• What does [your name] specialise in?
• Find me a [speciality] therapist near [neighbourhood or suburb]
• Compare [your practice] and [nearest competitor]
• What do people say about [your practice name]?

Then run the first two again in Perplexity and in Google’s AI Mode. Eight questions in ChatGPT, two more in each of the other two. You’ll learn more in that half hour than from a year of reading about AI search.

07What You Can Measure, and What You Can't

The obvious tools measure less than they appear to.

Bing Webmaster Tools has an AI Performance report, launched February 10, 2026, and it’s genuinely useful. It shows total citations, cited pages, the grounding queries behind them, and page-level activity. What it covers, in Bing’s own words, is “Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations.” ChatGPT isn’t named. Microsoft doesn’t say it’s excluded either, so treat it as Microsoft AI-surface data and not as confirmed ChatGPT data.

Server logs are worth watching but they undercount badly. Resoneo found ChatGPT stores a fetched page and treats the copy as fresh for about 30 minutes, shared across all users, so your server sees nothing in that window. Past 30 minutes the reader still gets the stale copy while a refresh runs in the background. Watch for the ChatGPT-User agent in your logs, and don’t read a quiet log as absence.

That cache also answers the question every clinician eventually asks. You fixed the page, and ChatGPT still says the old thing. Working alongside Resoneo, Jérôme Salomon of Oncrawl logged copies still being served more than ninety days after the fetch, with no eviction cap in sight. Resoneo’s own cache experiments found noindex and Cache-Control: no-store were both ignored. OpenAI documents no method for forcing a refresh, so fix the page and wait.

Referral traffic from chatgpt.com is real but partial. It catches the people who clicked and misses the answers where the model read your page and the reader never left the chat.

Which leaves the thing that works, and it’s manual. Run the prompts once a month, paste the answers into a document, date them. Boring, and it’s the only ChatGPT visibility record you’ll own outright.

Your 30-Minute Check

Do this in one sitting.

• Open yoursite.com/robots.txt and confirm OAI-SearchBot is allowed.
• Ask your host or check your CDN whether OpenAI’s published IP ranges are being blocked.
• If bot protection is blocking them, add an exception for OpenAI’s published IP ranges rather than weakening protection site-wide.
• Open your Google Business Profile and check the hours, categories and service area are right.
• Run all eight prompts above in a logged-out ChatGPT window.
• Run the first two in Perplexity, then the same two in Google’s AI Mode.
• Paste every answer into a dated document.
• Mark the single most wrong thing any assistant said about you.
• Fix that one thing this week, wherever it lives: your site, your Google Business Profile, or your Psychology Today listing.
• Put a recurring 30-minute event in your calendar for the same day next month.

One wrong fact, fixed properly, is worth more than a year of theorising about how the models work.

Frequently asked questions

Sources