Why AI Answers Show Your Brand Only Sometimes

By Ömer Cenk Tokgöz · Published:

Why AI Answers Show Your Brand Only Sometimes

Your brand is in a ChatGPT answer one day and gone the next. For B2B SaaS teams: why AI answers vary, how to measure it properly and what to do.

Short answer: AI answers are generated fresh every time. Even when the same question goes to the same engine twice, the answer can change. Add differences between engines, how the question is phrased, which pages the engine found at that moment and where the person asking is, and the variance grows. That is why a single screenshot proves neither a drop nor a gain. A proper measurement uses a fixed prompt set, repeated runs and a margin of error per engine.

B2B SaaS marketing teams often send us the same message: "Last week ChatGPT recommended us. Today I asked the same question and we are gone. What broke?" Usually nothing broke. This post explains why a brand appears in AI answers only some of the time, how to measure that properly and what you can change.

1. Same question, different answer

Language models generate an answer piece by piece and pick one of several possible continuations at each step. That choice involves some randomness. OpenAI's API documentation states that Chat Completions are non-deterministic by default, which means outputs may differ from request to request. In the ChatGPT or Gemini chat window, the user does not choose these settings.

In practice your brand is named in an answer every time, never, or some of the time. "Some of the time" is the most common case. In our 476-answer measurement (29 October 2026) we asked every question three times on every engine. Even the most-mentioned agency appeared in all three runs in 63 of 156 engine-question pairs. Showing up in one answer does not mean repeated visibility.

2. Each engine is its own source

ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode answer the same question with different models and different sources. Google says this openly about its own two products: AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.

We saw the same in our own measurement. The top two agencies were the same on all four engines, but third place changed by engine. So "Does AI recommend us?" has no single answer. It has one answer per engine.

3. How the question is phrased

Your buyer may ask "What is the best CRM?" or "Can you suggest an alternative to the well-known CRMs for a small B2B SaaS sales team?". The second question is longer, carries more context and can lead the engine to a different set of sources.

In our measurement, all four answers that recommended us first were on ChatGPT, and all four were sector questions (B2B SaaS, e-commerce, fintech). No general "best agency" question put us first. The same brand can rise to the top of the list or drop off it depending on how the question is built.

The lesson: your prompt set should contain the words your buyers actually use. Measuring only the category name ("CRM software") misses the questions closest to a purchase.

4. What the engine could find at that moment

Engines that search the web do so before writing the answer, then summarise the pages they found. Google says AI Overviews and AI Mode may use a "query fan-out" technique, splitting one question into several related searches. To appear as a supporting link in those answers, a page must be indexed and eligible to be shown in Google Search with a snippet.

That has two consequences. First, a page you just published or updated does not enter answers until the engine reads it again. Second, if the engine finds an old page of yours, it repeats the old information. We lived this on our own site: an engine carried prices from an old package card that had been left on our site into its answer. We describe it in our 476-answer measurement post.

5. Who is asking, and from where

Google says that even with personalisation turned off, it still uses your location, language, device type and current searches to improve results. With personalisation on, the order of results can change too. In chat-based engines, account settings, chat history and memory features can also affect the answer.

So a check run from the marketing team's own laptop is not neutral. The answer you see as someone who searches for your brand every day may differ from the one a buyer in another city sees. Run checks in a clean session, logged out, and note the location.

6. Which sources the engine relied on

When an engine recommends a brand, it usually starts from the pages it found. In our 29 October 2026 measurement, 375 of 476 answers cited at least one agency list page. Whoever wrote the list, the answer looked like it. Part of our own visibility also came from a list we wrote ourselves; we said so openly in that post.

For B2B SaaS this means: for your buyer's question, which comparison site, list article, forum or documentation does the engine cite? When that source changes, for example when a new list is published or an old one is updated, your visibility can change without you doing anything.

How to measure it properly

You cannot remove the variance, but you can measure it. Five rules are enough:

  1. Fix the prompt set. Ask the same questions in every measurement. If the questions change, you cannot compare two measurements.
  2. Repeat every prompt. One answer is a sample, not a measurement. Check whether the brand appears in all repeats of the same prompt.
  3. Read per engine. Write the answer count, mention rate and margin of error separately for each engine. A composite score is a useful summary, but it does not support a decision on its own.
  4. Know the margin of error. As our methodology page states, a scan of 200–300 answers has a margin of roughly ±4–5 points per engine. If the difference between two measurements is smaller than that band, we do not count it as a change.
  5. Log the source list. Which page was cited often explains why a rate moved better than the mention rate itself.

To set up the monitoring workflow step by step, see our AI visibility monitoring guide.

What you can do

  • Do not decide on one screenshot. Add the sales team's "we are gone today" message as a note for the next planned measurement.
  • Be present in the pages the engine cites. Appear with correct, current information on the comparison sites, list articles and directories that show up often in your measurement.
  • Tie your own pages to one truth. Price, package names, product features and company details should match everywhere on the site. An old page is still a source for the engine.
  • Cover your buyers' wording. For every question close to a purchase, write a page or section that answers in the first paragraph. How to split this work with your SEO team is covered in how SEO and GEO consulting work together.
  • Plan per engine. If you show up regularly on one engine and never on another, the first place to look is the pages that engine cites.

If you want us to run this measurement and fix loop for you, see our GEO services for B2B SaaS.

Limits of this post

The measurement examples here come from a single day and from agency questions. Variance in your SaaS category may be lower or higher. Engines also update often; behaviour that holds today may change in a few months. The only way to see it is to measure your own prompt set regularly.

Sources

Ömer Cenk Tokgöz

Ömer Cenk Tokgöz — Founder, Botfusions

Founder of Botfusions; focuses on Generative Engine Optimization (GEO) and AI / answer-engine visibility. Author of the AI Visibility methodology.

Author profile & areas of expertise → · Editorial standards →

Related Posts

All Posts