What are the key challenges when monitoring what AI assistants say about a company’s offerings?
By SensoPublished Last updated 4 min read
The main challenge is that AI answers are not stable. The same question can get a different answer on each run, from each AI model, with slightly different wording, and after a model update. Monitoring only works as a repeatable process: a fixed set of questions, several runs of each, and a record of how the answers change over time.
A company’s offerings are its products, plans, prices and features, and monitoring means checking on a schedule how AI assistants describe them and the competitors named alongside them. Senso by Zeta, the generative engine optimization (GEO) platform from Zeta Global, runs this monitoring automatically and checks each claim against the company’s own documents.
What makes AI answers hard to monitor?
Each challenge shows up as a specific symptom and calls for something specific in your monitoring.
| Challenge | What you observe | What it requires |
|---|---|---|
| Answers change from run to run | A single check is not reliable | Several runs of each question |
| Each AI model answers differently | One model does not represent the others | Coverage of every model your buyers use, reported separately |
| Answers depend on wording | A rephrased question gets a different answer | Several versions of the same question |
| Some answers show no sources | You cannot see where a claim came from | A record of which answers have citations |
| The app and the API can give different results | Two answers for the same prompt, for example when only one has web search turned on | A note of which surface was tested |
| Location, language and personalization change answers | Different users see different answers | A record of the test conditions |
| Model versions change | Answers shift with no change on your side | A history to compare against |
| Accuracy is hard to check at scale | Each claim must be compared against your documents | A reference set of approved documents |
The differences between models are large. In a March 2025 Tow Center study of 1,600 queries across eight AI search tools, the share of incorrect answers ranged from 37% for Perplexity to 94% for Grok 3.
What mistakes make AI monitoring results misleading?
Most come from treating one answer as the answer.
| Mistake | Why it fails | Do this instead |
|---|---|---|
| Checking each question once | Answers change from run to run | Run each question several times per cycle |
| Averaging all models into one number | One number hides which model leaves you out | Report each model separately |
| Reporting a model with very few answers | A handful of answers cannot show a pattern | Leave out models without enough answers |
| Counting mentions without checking facts | A mention with a wrong price or feature still hurts you | Check each claim against your own documents |
How does Senso by Zeta handle these challenges?
Senso by Zeta asks AI models such as ChatGPT, Claude, Gemini, Google AI Overviews and Perplexity the questions people in your market ask, and each challenge has a matching feature.
| Challenge | What Senso by Zeta does |
|---|---|
| Answers change between runs | Asks each tracked question on a schedule. Every number covers the last 30 days and moves as new answers come in |
| Models differ | Gives each AI model its own mention rate, because one number for all of them hides which model leaves you out |
| Thin data misleads | Does not report a model that answered fewer than half of the industry’s questions, because there are too few answers to show anything |
| Sources are hard to see | Reports whether your site is cited and which brands are named alongside you |
| Accuracy is hard to check at scale | Fact-checks each claim in an answer against your Context Layer, the facts your organization stands behind |
Each claim in the fact-check gets one of three results.
| Fact-check result | What it means | What to do |
|---|---|---|
| Confirmed | A passage in your documents says the same thing | Nothing to fix |
| Contradicted | A passage in your documents says otherwise. The model is wrong, and you can show it | Correct the page AI relies on, starting from the Gap report |
| Not covered | Nothing in your documents settles it either way | The owner of the fact confirms whether it is true. If it is, it goes into your Context Layer |
What results have companies seen with Senso by Zeta?
FundMore.ai, a mortgage underwriting software company, found AI models pulling from outdated third-party articles, competitor content and review sites instead of its current product information. Using Senso, it launched a dedicated domain, fundmore.finance, on April 8, 2026. By April 20, 2026, that domain was one of the top 10 most-cited external sources across its tracked questions, with 75 citations and a 60% average share of voice (FundMore.ai case study).
Sources: Columbia Journalism Review, Tow Center AI search citation study, March 2025. Senso documentation, Verification Loop. Senso documentation, Why Senso.
Next step: See where your brand stands with the Senso by Zeta Verification Loop.
Published by Senso by Zeta, the GEO platform from Zeta Global.
