How to Check Whether Your Company Appears in AI Answers

The method is simple. Pick one question that a customer would realistically type, and enter the same sentence five times in a new chat on each engine. If you write down which answer your company's name appeared in and which sources that answer linked to, you can measure again under the same conditions next month and compare.

The most common mistake in this measurement is deciding whether your company appears or not based on a single screen you asked once.

Fix the Conditions First

Even the same question produces different companies and sources in the answer depending on the service, the timing, and what came before in the conversation. This is also why GEO Times records one question repeated across engines (editorial principles).

That is why you have to fix the conditions to place results from two points in time side by side. Keep the question sentence exactly the same on every round, and start a new chat each time. Set ChatGPT to logged-out or Temporary Chat, and for Gemini and Claude, which require login, decide on a condition you don't normally use and keep it for all five rounds.

Do all the measuring together, on the same day and at the same time of day. If you ask from your own account after a long history of work conversations has built up, you will see an answer mixed with that history.

When choosing a question, write one that leaves out your company's name first. Duane Forrester wrote in a Search Engine Journal article that asking a model to explain a named company is the least informative kind of test.

It is the question without your company's name that determines whether your company gets included in the answer. This includes recommendation questions that include only a region and industry, and comparison questions that mention only one competitor.

Record the Model Name Too

Google AI Mode added Gemini 3.7 Flash to its model menu on August 14, and 3.8 Flash appeared in the paid subscriber menu on September 2. Right after launch, a problem was reported where 3.8 Flash answers carried almost no source links, and it was reversed within a day (Google adds Gemini 3.8 Flash to AI Mode).

Search Engine Journal, reporting this news, wrote to record the model name every time you test. A difference between a value measured with 3.7 Flash in mid-August and one measured with 3.8 Flash in September could be the result of the model changing, not your content.

The free version of AI Mode has no model selection menu, so the answer comes from the default model. In your table, note whether the answer came from a model chosen on a paid account or from the default model.

What to Include in Your Log Table

One spreadsheet is enough. Fill in the eight columns below, one row per round.

Column	What to enter
Measurement date	Date and time of day
Engine and model	Service name, name of the model chosen
Question	The exact sentence you typed
Round	1 through 5
Your company's appearance	Whether it appeared, and at what point in the answer
Companies mentioned alongside	Names exactly as written in the answer
Source domain	Domain of the page the answer linked to
Source type	Whether it's your own site or a third-party site

Splitting the table this way lets you count your company's name appearing in the text separately from your page being attached as a source. The two do not move together. Sometimes the name appears but the link goes to a news outlet or blog, and sometimes only your page ends up in the source list with no name mentioned.

Write down every company name that appears alongside yours, without leaving any out. If the same three companies show up across all five rounds, the table itself shows who you are being compared against in that answer.

Recording source domains gives you evidence for choosing which page to work on next. In 7,387 citations that the US B2B content company Ten Speed gathered using 170 prompts, product pages accounted for the most at 24.1%, comparison pages and listicles each came in around 13%, and directory profiles like G2 accounted for 7.2% (Product Pages Get 24.1% of AI Citations, Reddit 4.2%). These numbers carry the conditions of B2B SaaS clients, English-language questions, and a single point in time, so for a domestic industry, the right approach is to count the source domains in your own table.

Check Whether the Description Given Is Accurate

Counting only whether you appear misses something. In a September 3 Search Engine Journal article, Duane Forrester wrote that when there isn't much material about a company on the web, instead of leaving the blank empty, a model pulls in the closest material it can find and presents it with the same confidence (When AI Has Nothing On Your Company, It Describes Someone Else).

The four patterns he described are: attaching a competitor's description to your company, stating outdated information in the present tense, treating a single piece of evidence with the confidence of multiple sources, and attaching your company's name to a generic industry statement. All four read plausibly enough that a quick skim lets them slide by. An answer that is really just a generic industry statement with your company's name attached is especially easy to pass over, since nothing in it is wrong as a description of the industry.

Forrester wrote that no amount of checking your own site pages will reveal this kind of mismatch. Whether an answer contains a sentence you never wrote is something you can only see by opening the answer itself.

Add one more column to your table and copy down, word for word, any sentence in the answer's description that isn't true. Out of the five times your company's name appears, how many times the description was correct is a value as important as the count of appearances itself.

When to Measure Again, and What Not to Do

Measure again under the same conditions one to two weeks later. It is not strange in itself for the result to change, so read it as a count out of five.

Save a screenshot of the answer for each round. If only numbers are left in the table, you won't be able to trace back later what sentences actually appeared.

Once the table is filled in for one question, add two or three more the same way. Expanding in the order a customer would actually ask makes it easier to later find where the answer changed.

Decide on what not to do as well. That means drawing a conclusion from a single answer, slightly changing the question wording each round, measuring from an account with an existing conversation history, and asking an AI to rank companies and using that score as your measurement.