Written by Tristan Michel
The right competitor list starts with businesses that can win the same customer and the same job. If an Austin homeowner needs a whole-home repipe, the useful rivals are companies that can quote that work, not every plumber in Texas or every website ranking for a broad plumbing term.
The audit also needs a second list: businesses that ChatGPT and other AI tools actually name for the customer questions being tested. The two lists won't match. Specialists, premium alternatives and unexpected rivals can enter the answer even when the owner never put them into a conventional SEO report.
One competitor list will miss part of the market
A local owner usually knows who quotes against them, advertises in the same area or comes up in sales conversations. Those businesses belong in the audit because they represent real commercial competition, even when an AI answer leaves them out on a particular run.
AI answers can introduce a different group. The wording of the job can change which businesses are suitable enough to mention. OpenAI says ChatGPT Search may rewrite a question into more targeted web searches and may use location information for local results. It does not publish a fixed formula for local recommendations or guarantee placement.
A small AnswerPrism check shows why the job matters. Two fresh ChatGPT Search questions asked for Austin plumbers. One concerned an emergency burst pipe in an older home that night. The other concerned a planned whole-home repipe in a 1960s house. Only one company appeared in both recommendation sets. Three of the first four names changed with the job, and the repipe answer introduced another candidate later in the response.
The sample is too small to describe the Austin plumbing market or explain why any company appeared. It does show the risk of treating one broad competitor list as though it covers every customer decision.
Build the list in two passes
Build the list in two passes. You need two separate lists. The planned list is a stable set drawn from actual, hard evidence of who sells the same job today. You keep it fixed throughout the measurement period.
The observed list begins empty. Add a name when it returns across the questions or collection windows that matter, and keep one-off appearances in the working notes until there is enough reason to treat the business as a recurring rival.
For the planned list, start with sales records, estimates, customer comments and the businesses your team encounters in the market. Ask a plain question: could this company credibly win the work the customer described?
A residential roofer does not need every construction company in the city. It may need another residential roofer, a storm-damage specialist and a premium contractor that serves the same properties. A marketplace or directory can occupy the customer's attention, but it should be labelled as a platform rather than made to look like a roofing company.
Useful groups include:
- Direct competitors that sell the same service in the same working area
- Specialists that become more relevant when the job includes a particular material, property or deadline
- Premium or budget alternatives a real customer might choose instead
- Businesses that recur in the collected AI answers
- Directories, marketplaces or publishers that route customers toward other providers
Leave irrelevant names out
A longer list does not make the comparison more honest. A famous national company may be irrelevant if it never serves the location or the job. A nearby company may be irrelevant if it handles a different customer, property type or price range. A business seen once may simply reflect ordinary answer variation.
Do not add a company merely because it ranks for a broad keyword. Search competition can help find candidates, but the final test is commercial: can it take the same work under the conditions in the question?
The same boundary applies to an AI-generated name. Check the business before treating it as a rival. It may be outside the service area, permanently closed, misdescribed or unable to do the job. The answer is an observation, not a verified market map.
Keep the benchmark still while the answers move
The planned list should remain stable for the comparison period. Otherwise, a business can appear to gain ground simply because stronger competitors were removed from the report. Keep an observation log beside it for new names.
Don't change the benchmark to chase an appearance on the current sample. Similarly, don't add every one-off name you see in the current run to the stable benchmark. At the same time, copy any new names you see in the current run into the observed list, to be reconsidered in future runs. Remember: one run is too small a sample to change your planned list. Keep your benchmark fixed and your observed list growing as you go.
A recent preregistered study found low agreement among four AI systems and substantial churn across repeated local food-and-drink questions. It covered two Bali markets, not American service businesses, so it cannot predict how often a plumber or dentist will appear. It supports the simpler measurement rule: record the conditions and do not rebuild the competitor set around one answer.
Compare reasons, not only names
Compare reasons, not only names. Record each competitor appearance not just as a name, but using all the details you collected about the question and context for each answer.
Keep the exact customer question, product and search mode, location wording, businesses named, whether each one was merely mentioned or actually recommended, the reason given and the visible sources. Open the sources before relying on them. A citation may support background information without supporting the recommendation itself.
Look at the recorded cases to see if you can distinguish four kinds of outcome: a real fit gap that needs fixing, a documentation gap that needs fixing, a factual error that needs fixing, or ordinary answer variation and nothing more.
A fit gap may be legitimate: the competitor offers emergency service and you do not. A documentation gap means the company may fit, but its website and public profiles do not make that clear. A factual error needs correction at the source carrying it. Ordinary variation may require no website change at all. This is also why a competitor recommendation should be diagnosed before its copy is imitated.
Use a list you can defend
Pick a small, defensible planned list. Pick a realistic set of questions that someone could really ask and expect an answer. Run them all under the same recorded conditions. Add recurring names from your own observed list, without pretending the list is yet complete or permanent.
Review the planned list when the service mix, territory or customer changes, not every time an answer moves. The purpose is to see who wins the decisions the business actually cares about and what evidence separates them.