Waarom veranderen AI-antwoorden in de loop van de tijd?

You run the same prompt on a Monday, paste it again two weeks later, and the answer feels like it came from a different system. The shift can be small (different bullets, different examples) or big (different conclusions, different sources, different tone). If you are doing QA, policy review, knowledge base validation, or brand visibility audits, this “week-to-week drift” is not a curiosity—it is a reproducibility problem.

What “drift” looks like in real AI use

AI answers rarely change for just one reason. Drift is usually the result of several moving parts: model updates, retrieval changes, safety filters, and randomness in generation.

It helps to describe the symptoms precisely before hunting for causes. Different symptoms point to different fixes.

• Content drift: facts, examples, or recommended steps change.
• Format drift: headings, tables, JSON fields, or length constraints stop being followed.
• Source drift: the same question gets citations from different domains or none at all.
• Stance drift: the assistant becomes more cautious, more confident, or changes its conclusion.
• Entity drift: brand names, product capabilities, or definitions get mixed up.

Once you name the drift type, you can isolate whether you are seeing “the model changed,” “the web changed,” or “your testing conditions changed.”

Why AI answers drift over time

Most people assume prompt wording is the main driver. It matters, yet it is rarely the largest driver in repeated tests. The bigger drivers live outside the prompt: the model version, the retrieval layer, the environment, and the sampling settings.

1) The model itself gets updated (quietly and often)

Many AI products ship continuous improvements. A “model update” can mean a new base model, new fine-tuning, new system prompts, new safety rules, or a new way of routing your request.

Even if the UI looks unchanged, your prompt may be handled by a different internal stack than last month. That can change style, refusal behavior, reasoning depth, and what the model treats as “allowed.”

• Some updates change how strictly constraints are followed (format, tone, length).
• Some updates change default uncertainty behavior (more hedging, fewer specifics).
• Some updates change what the assistant prioritizes (citations vs synthesis, brevity vs completeness).

2) Randomness is part of how text generation works

Even with the same prompt, many systems do not produce a single deterministic “correct” answer. They sample from a distribution of likely next tokens.

Small differences early in a response compound. A different first sentence can lead to a different outline, which leads to different examples, which leads to different conclusions.

If you are auditing drift, treat one run as anecdotal. Do controlled reruns and measure the spread, like you would in any stochastic system.

3) Retrieval results change as the web and the index change

When an assistant uses live web retrieval, you are not only testing the language model. You are testing an entire pipeline: query rewriting, document selection, ranking, extraction, and citation formatting.

Those layers shift because the web shifts. Pages get updated, deleted, paywalled, redirected, or pushed down by new competitors.

• Index freshness: what the assistant can “see” depends on what its underlying index has crawled recently.
• Ranking shifts: a small algorithm change can swap which sources are selected.
• Snippet selection: even if the same page is retrieved, a different passage might be extracted.

If you work with SearchGPT-style experiences, the retrieval index matters a lot. One practical implication is that a site can be visible in one ecosystem and effectively invisible in another, which shows up as source drift and missing citations. The mechanics of cross-tool prompt testing, including how to log retrieval differences, are covered in hoe je prompts kunt testen in ChatGPT, Gemini en Perplexity.

4) Safety, compliance, and policy guardrails evolve

Refusals and “I can’t help with that” answers can change even when the prompt stays constant. That does not always mean the model “got worse.” It may mean policy interpretation changed, or the tool became stricter in borderline domains.

Policy drift is common in areas like medical, legal, finance, cyber security, and sensitive personal data. It can show up as:

• More disclaimers and fewer actionable steps
• More requests for clarification
• More conservative recommendations

If this is the drift you see, the fix is rarely “rewrite the prompt.” The fix is to narrow scope, specify the intended use, and remove risky ambiguity so the assistant can answer safely without refusing.

5) Your session context changes the answer, even if you think it doesn’t

Many chat tools use conversation history as part of the input. If you test in a reused thread, you are not running the same prompt. You are running “prompt + hidden context.”

Context can include prior messages, tool state, custom instructions, memory features, or even your selected language and region. It can affect what the assistant assumes, which examples it picks, and which risks it flags.

• New chat vs ongoing chat
• Logged in vs logged out
• Region and language settings
• Browsing/sources mode toggles

6) Tooling around the model changes (connectors, search modes, citations)

Even when the base model stays similar, product teams update the “wrapper” constantly. That wrapper can include browsing modes, citation formatting, connectors to internal documents, or different routing for certain prompt types.

This is why two experiences that both say “ChatGPT” or “Gemini” can behave differently depending on the mode you used. In practice, this is one reason brand and citation audits need a stable protocol and a clear log of mode flags.

How to diagnose what caused the drift

Diagnosis is about logging the right variables, not collecting every variable. You want to reduce the number of plausible explanations until one remains.

Use a simple drift triage table

This table exists to route you to the right next check without guessing.

Run controlled reruns and log the spread

A practical baseline is three runs, each in a fresh session, within the same hour. If the spread is wide, you are measuring randomness or ambiguity, not “a stable capability.”

For teams that need a repeatable QA workflow, the rubric-and-log approach from testopdrachten voor ChatGPT, Gemini en Perplexity is a good operational standard: it forces you to separate input, system behavior, and output.

Track metadata that makes changes explainable

You do not need a heavy tool to do this. A spreadsheet is enough if it captures the fields that explain drift.

• Date/time window
• Tool and mode (browsing/sources on or off)
• Regio/taal
• New chat or existing chat
• Prompt ID and exact text
• Output and citations
• Score (accuracy, completeness, format compliance)

How to reduce drift when you need reproducible outputs

You can’t freeze the entire ecosystem, yet you can make your tests stable enough that differences are interpretable. The goal is not “identical text forever.” The goal is “stable evaluation and auditability.”

Make the prompt less ambiguous, not longer

Ambiguity creates variance. Tight scope reduces variance.

• Geef de doelgroep en de context aan (functie, branche, beperkingen).
• Define what is out of scope (“Do not cover X”)
• Require an output shape (bullets, table, fixed headings)
• Ask for assumptions explicitly (“State assumptions in 3 bullets”)

Standardize your environment

Use new chats for baseline tests. Keep region, language, and mode flags consistent within a cycle. Avoid mixing logged-in and logged-out tests unless that difference is the point.

Separate “evaluation prompts” from “production prompts”

Prompts used for measurement should be stable and boring. Prompts used for daily work can evolve.

If you blur the two, you lose the ability to say whether outcomes changed because the tool changed or because your team changed its prompting habits.

Use authoritative reference points for definitions

When teams argue about whether an answer “changed,” disagreements often start with different definitions of the same term. A neutral baseline can prevent that.

For example, if you need a shared definition of generative AI across your audit docs, Wikipedia’s overview is a common reference: Generatieve kunstmatige intelligentie.

What drift means for brand visibility and audit trails

If your use case is brand monitoring—whether assistants mention you, cite you, or describe you correctly—drift is not noise. It is part of the system you are measuring.

This is why visibility work needs metrics that survive week-to-week output changes. The measurement framing in vertrouwenssignalen die de kans op citatie door AI vergroten can help you focus on what stays stable: retrievability, attribution, and extractable passages.

If you want a lightweight, repeatable way to track drift, test prompts, and connect findings to a publishing and internal linking plan, Authora can help you set up a structured workflow that supports reproducible AI audits over time.