Guides › How to Get ChatGPT to Recommend Your Business
How to Get ChatGPT to Recommend Your Business
There is no setting that makes ChatGPT, or any assistant, recommend a specific business: no submission form, no paid placement, no ranking algorithm to game. What you can control is whether the assistant gets your business right when someone asks about it, and that is a measurable, fixable problem. This guide covers what we've actually tested, in order of what moved the number.
Why assistants misread businesses in the first place
Three separate failure modes showed up repeatedly when we sent real AI models at real business websites and at real business names:
The agent never reaches the right company. Asked about nine real companies by name, with no tools and no ability to search, models gave the correct URL in only 12 of 54 answers, 22% (Note 14). The good news buried in that number: models were not confabulating. 39 of 54 answers explicitly admitted low confidence, and confidence was well calibrated: answers given with confidence at or above 0.5 were correct 85% of the time; below 0.5, correct only 2% of the time. When a model doesn't know your business, it usually says so. The failure is not memory, it's retrieval, and asked directly, models will name exactly who they confuse you with (a competitor, a similarly-named company, sometimes a company on a different continent entirely).
The agent reaches you and stops after one page. Asked a simple, casual question ("find out what this is"), six of six models across two real business sites read only the homepage and stopped voluntarily (Note 01). Everything on your services pages, case studies, and blog archive may as well not exist for that query. The same models will read seven to fifteen pages when the question demands it (Note 02). Depth is task-bound, not a fixed behavior.
The agent reaches you, reads you correctly, and still can't answer the question that matters. On eight real client sites across 48 recorded agent visits, not a single one produced a price: 100%, and it reproduced across three independent passes (Notes 12, 13). Comprehension was fine; seven of eight sites had every model agreeing on what the company does. The commercial facts were simply missing from where the agent was reading.
The fixes, in order of measured impact
1. Put the answer on the page first
This is the fix with the largest measured effect. Taking a fact that was missing and placing it directly in the first few hundred words of the homepage, not linked, not in a separate file, took recovery on a real client page from 30% to 78%, with agents needing an average of 1.1 fetches instead of hunting through the site (Note 41). A single inlined price took a control group's price-recovery rate from 0 of 10 to 10 of 10 (Note 19). This works because it removes the step that fails most often: making an agent find and then use a second document, rather than reading the answer where it already is.
2. Answer the buyer's actual question, not the spec sheet
A fact only earns its place if it's translated into what a buyer is really asking. "IP65" is a specification; "suited to outdoor and public-space installations" is an answer. On one real client rewrite, a block built from what was easy to find rather than what buyers ask about scored no better than nothing on the questions that mattered, because it answered questions nobody was asking (Note 38).
3. Fix your own name collision
If a model has ever confused your business with a different company of a similar name, that is diagnosable directly: ask a model, with no search tools, what it believes about your company name, and it will often name the collision unprompted. This costs nothing to check and tells you exactly what disambiguation work is worth doing.
What doesn't work
- Publishing llms.txt alone. Without an effective, visible pointer
to it, agents did not find it: 0 of 18 test visits reached an unpointed
llms.txton a real site (Note 15). Even when they did fetch a file containing a real answer, they often didn't use it: 6 of 10 runs fetched a file with a real price in it, and none of those ten reported the price (Note 19). See our full llms.txt guide for what the file is actually good for. - robots.txt. Not one of six models fetched
robots.txtin any of our tests, on either of two sites, while our own server logs show crawlers like GPTBot and ClaudeBot requesting it first, every time. Agents answering a buyer's question and crawlers indexing your site are different populations, and robots.txt reaches only the second one (Note 04). - Meta tags.
<meta name="llms-txt">scored worse than doing nothing at all in our signposting tests, 0 of 4, because browsing tools receive extracted visible text, and<head>markup does not survive that extraction. What the intuition says is the strongest position turned out to be the weakest (Note 06).
Realistic expectations
None of this is a guarantee that an assistant recommends you over a competitor for a given query. That outcome depends on the question asked, the competing businesses in range, and factors outside any single site's control. What is measurable, and what we sell, is narrower and honest: whether an agent that reaches your site comes away with the facts right instead of missing, wrong, or attributed to someone else. That is a testable claim, and we test it before and after every engagement.
Find out where you stand
Run a free audit and see what an AI agent currently gets right and wrong about your business, or see the full fix.