In brief

Reliable AI monitoring measures brand mentions in ChatGPT, Gemini, Claude and Perplexity while controlling for the systematic biases of the models. Four rules are decisive: ask brand comparisons in both orders, phrase prompts without built-in judgements, disclose the limits of the measurement and check the results against an independent industry report. Loud & Lexis applies these rules in its AI Visibility Monitoring.

Last updated: 12 November 2026

You set up AI monitoring correctly by asking brand comparisons in both orders, phrasing prompts without built-in judgements, disclosing the limits of the measurement and reading the results against an independent industry report. For brands that build their visibility with the Fishnet Strategy (in German), monitoring is the measurement strand of the content network. Loud & Lexis measures this visibility continuously across ChatGPT, Gemini, Claude and Perplexity with its AI Visibility Monitoring (in German). It pays off commercially because in 2024 Gartner predicted that traditional search engine volume will drop by 25 percent by 2026 (MediaPost, 2024).

At a glance

Why do brand comparisons need to be asked in both orders?

Brand comparisons need to be asked in both orders because, when AI rates two brands as equally strong, the brand named first wins 63.3 percent of head-to-head comparisons. This figure comes from the Loud & Lexis whitepaper “First come, first recommended?” (Henrich, Park and Peschke, October 2026). The underlying position bias is independently documented: a systematic study (Judging the Judges, arXiv 2025) shows that AI models acting as judges favour answers based on their position in the prompt instead of judging on content alone.

In its AI Visibility Monitoring, Loud & Lexis asks every brand comparison in both orders and only counts a lead if it holds in both runs. Anyone who tests only one order partly measures the position of the name, not the strength of the brand. The methodological basis is set out in the analysis of position bias.

How does the wording of a prompt distort monitoring?

The wording of a prompt distorts monitoring as soon as it contains a judgement, a brand name or an assumption that the model adopts. A study of 14 AI judges (When Wording Steers the Evaluation, arXiv 2026) shows that the framing of a question alone shifts a language model’s answer considerably. If a user opinion is included in the prompt, the rate of agreement with factually incorrect assumptions rises to 63.7 percent on average (When Truth Is Overridden, arXiv 2025). AI systems are therefore easy to steer. This is exactly what makes neutral monitoring difficult: anyone who writes the bias into the prompt distorts the measured visibility. Loud & Lexis therefore builds its tracking prompts from a neutral core question plus controlled variants instead of a single, judgemental wording.

Which user contexts can AI monitoring not capture?

AI monitoring cannot capture two user contexts: the actual course of the customer journey in the chat and the information a user has stored in their AI account. The digital customer journey is barely documented by data, because nobody sees users’ real chat histories; an analysis of publicly shared chats (WebFX, 2025) counts only 1.7 messages per conversation on average and only approximates multi-step research across the five research levels of an agentic customer journey (in German). Then there is the account context: according to OpenAI’s documentation (Memory and new controls for ChatGPT, 2024), ChatGPT draws on saved memories and previous conversations in new answers and tailors them to the individual user. Loud & Lexis therefore measures a controlled model of reality and openly states this limit in its monitoring.

How can monitoring reveal where AI misjudges reality?

Where AI misjudges reality becomes apparent when monitoring results are compared with an independent industry report that collects data on competitors, the market, determinants and personas on its own. For this purpose, Loud & Lexis has built the industry report Krynoscope together with Opascope. Krynoscope takes all the bias factors mentioned into account and creates such a report around the company. If an AI answer deviates from the report, this reveals a misunderstanding on the part of the AI and, at the same time, where strands are missing in the content network.

Bias factor Countermeasure Implementation in Loud & Lexis monitoring
Position bias Ask both orders A lead only counts if it holds both ways
Framing bias Neutral core question, controlled variants Prompts without judgements
Account context Neutral measurement setup Limit openly stated

Loud & Lexis covers which factors AI assistants actually weigh in comparisons in the article on determinants. How AI visibility can be measured overall is shown in the overview Measuring AI visibility (in German).

Selection criteria for reliable AI monitoring

Reliable AI monitoring can be recognised by its methodology, not by its range of features. The following criteria separate meaningful measurement from false precision:

Frequently asked questions

Meaningful AI monitoring needs a persona-based prompt set across all five research levels that runs in parallel against the four dominant models. Besides the number of prompts, repetition is what matters most: an empirical study (Don’t Measure Once, arXiv 2026) recommends at least seven runs per prompt and day, because measured visibility fluctuates between runs.

AI monitoring should use a neutral, non-personalised account. Logged-in accounts distort the result, because saved memories and previous conversations tailor the answers to the individual user, as the OpenAI documentation mentioned above shows. Loud & Lexis therefore measures without personal account context.

AI monitoring only delivers reliable figures through repeated measurement, because AI systems are not deterministic and the same question can produce different answers in different runs (Equal Experts, 2025). At Loud & Lexis, the first reliable trends typically emerge after several weeks of continuous measurement.

AI monitoring can be set up with free tools for initial spot checks, but not for reliable ongoing measurement. Repeated queries across several models, both orders and several weeks usually require a paid subscription. Loud & Lexis uses Promptwatch as its measurement system for this and discloses the ongoing costs to the client transparently.

Their own AI monitoring does not pay off for brands without a defined competitive and persona environment and without content that can respond to visibility. Without this basis for comparison, the monitoring lacks the industry report against which the AI answers are read. The measurement then has no practical consequence.

Strategy call

How visible is your brand
to ChatGPT & Co.?

Free 30-minute strategy call with an initial assessment of your AI visibility.

Book a strategy call →