Why AI assistants recommend your competitor, and how to find out
In short
An AI assistant recommends the brand it can describe most confidently from the pages it reads, and most of those pages are not yours. To find out why it picked a rival, ask it the question, keep the whole answer and its links, and sort the reason. The usual reasons are a third-party list of tools, a comparison page, pricing it could read, or a directory.
The short answer
Assistants with web search build an answer from pages they find for the question. If those pages name your competitor and not you, the answer usually does too.
Your own site is one page among many, and often not one of them. The pages an assistant finds for "best tool for X" are usually lists, guides and reviews written by other people.
How an assistant builds a recommendation
The steps differ by assistant, but the shape is the same when web search is on.
- It reads the question and decides what to search for.
- It fetches a handful of pages that match.
- It writes an answer that names the brands those pages agree on.
- It may link some of those pages as sources.
Without web search, it answers from what it learned in training. That is harder to trace, because there are no links to follow.
The common reasons a rival is picked
When you open the pages a losing answer cited, the reason usually falls into one of a few kinds.
| Reason | What you see |
|---|---|
| A list of tools | The answer cites a "best tools" article that names them and not you |
| A comparison page | The answer leans on a page comparing them with others |
| Readable pricing | The answer quotes their price and not yours |
| A directory | The answer comes from a listing or marketplace |
| No clear source | The answer cites nothing useful, or nothing at all |
How to find out, by hand
You can do this without any tool. Ten questions on four assistants is forty answers to read.
- Write ten questions a buyer would ask before they know your name. Leave your brand out of every one.
- Ask each question in each assistant you care about, with web search on, in a fresh chat with no history.
- Save every answer in full, with its links. A screenshot loses the links, so copy the text.
- For each answer, note who was recommended, and whether you were.
- Open the pages each losing answer cited, and sort the loss into one of the reasons above.
- Ask the same questions again a few days later. Answers change on their own, and a loss that appears once may not be a loss.
Keep the exact sentence each answer wrote about each brand. A note such as "it preferred the rival" cannot be checked later. The sentence can.
A worked example
The numbers in this example are made up to show the method. They are not a measurement of any real company.
Say you sell accounting software for small shops. You ask ten questions on four assistants and get 40 answers. One rival is named in 18 of them. You are named in 2.
You open the pages cited by the 18 answers that named the rival. Most of them lead to the same two "best accounting software" lists, and you are on neither. A few lead to the rival's own comparison page. The rest cite nothing you can act on.
Now you have a plan in order. Ask to be added to the two lists. Write a fair comparison page of your own. Then ask the same ten questions again and compare.
The pages an answer cites are the evidence. The answer only tells you the verdict.
What can change each reason
| Reason | What could change it |
|---|---|
| A list of tools | Getting onto that list, and onto the next ones written |
| A comparison page | A fair comparison page of your own |
| Readable pricing | Prices in plain text on your own site |
| A directory | A listing if you can get one. Often nothing a page can do |
| No clear source | A clear page on the exact question, then ask again |
Agentcites does this sorting for every lost question, on every scan, and says which losses a page of yours could win. The why you lost page shows how.
Why the answer changes when you ask again
Assistants do not give the same answer every time. Ask twice and the sources, and sometimes the brands, can differ.
A 2026 preprint, CITECHOICE, found that about one in seven regenerations of the same answer flipped whether a given source was cited. It has not been peer reviewed, but it is a good reason to ask more than once before you act.
What this cannot tell you
- It cannot prove a page caused the answer. A cited page and a recommendation appear together. That is a strong clue, not proof.
- Being cited is not being recommended. An assistant can read your page and still name someone else.
- One round of asking is a single reading. Repeat it before you act on a small difference.
Questions
Why does ChatGPT recommend my competitor and not me?
Usually because the pages it reads for that question name your competitor and not you. Those are most often third-party lists and comparison pages, not your own site.
How do I find out which sources an AI assistant used?
Ask with web search on and keep the links in the answer. Some assistants link most of their claims. Others give fewer links, or send them through their own redirect address, so the exact page is harder to see.
Can I make an assistant recommend me?
Nobody can promise that. You can change the pages it reads: get onto the lists it cites, publish a fair comparison page, and put your prices in plain text. Then measure again.
Why does my competitor show up even when the answer cites no sources?
The assistant may be answering from what it learned in training. A brand written about often and clearly before the training ended is easier for it to name. That changes slowly, so work on the pages it reads with web search first.