Why might a model start pulling from different sources over time?

Most models do not pull from one fixed source. They query a retrieval layer that changes as raw sources are ingested, compiled, re-ranked, and removed. When that layer moves, the same question can surface a different citation over time. That is source drift, and it matters any time answers need to stay grounded and provable.

Quick answer

The model usually starts pulling from different sources because the surrounding system changed, not because the question suddenly changed.

Common triggers include source updates, index rebuilds, prompt changes, model version updates, permission changes, and ranking rules that favor freshness or authority.

In retrieval-augmented generation, the generator writes the response, but the retriever decides which raw sources enter the context. If the retriever changes, the citation path changes with it.

Common causes of source drift

Cause	What changes	What you see
Raw source updates	Policies, pages, or docs are edited	The model cites newer material
Re-indexing or re-compiling	The compiled knowledge base is rebuilt	A different source ranks first
Prompt or instruction changes	The query is interpreted differently	The model asks for different evidence
Model version updates	The base model or tool routing changes	Source selection shifts after a release
Access control changes	A source becomes hidden or blocked	The model falls back to another source
Freshness bias	Recent sources outrank older ones	Older citations disappear
Query ambiguity	Small wording changes alter intent	Different source clusters appear

Why a model starts pulling from different sources over time

1. The raw sources changed

If a policy page, product page, or help article changed, the model may pull from the updated version. If the old page was retired or redirected, the model has less reason to use it.

This is common when teams update content without versioning the source history. The answer still looks normal, but the evidence behind it has changed.

2. The compiled knowledge base was rebuilt

In many enterprise systems, raw sources are ingested and compiled into a governed knowledge base. When that compilation runs again, the source ranking can shift.

New content may enter the top results. Old content may drop out if it no longer matches the retrieval rules. The model is not inventing a new source. It is using a different retrieval path.

3. The prompt or orchestration changed

A small prompt edit can change which evidence the model asks for. Tool routing can also change.

If the system now prefers product docs over policy docs, or internal sources over public sources, the answer path changes even if the user query does not. That is one of the fastest ways source drift appears.

4. The base model or vendor stack changed

Vendor updates can change how the system interprets intent, ranks evidence, or calls tools. If the model version changes, source selection can change with it.

That is one reason teams see different citations after a release even when their raw sources stay the same. The generator is only one part of the stack.

5. Access or permissions changed

A source may still exist, but the model may no longer reach it. Access rules, region restrictions, login states, or compliance filters can block a source from retrieval.

When that happens, the system falls back to another source that is easier to reach. The answer may still be valid, but the citation trail is different.

6. Freshness rules changed

Many systems rank newer material higher. That is useful when the latest policy or pricing matters. It becomes a problem when freshness outranks authority.

The model may move from a stable source to a newer but weaker one. That creates inconsistency over time, especially in regulated workflows.

7. The question is not as stable as it looks

Two prompts that look similar to a person can be very different to a retriever. A single word can shift the intent from pricing to policy, or from internal guidance to public messaging.

That changes the source set. If the prompt drifts, the citations will drift with it.

Why this matters for teams

When source behavior changes, the risk is not only bad answers. The bigger problem is that the organization cannot prove why the answer changed.

For marketing teams, that affects AI Visibility and brand narrative control.

For compliance teams, that affects citation accuracy and auditability.

For security and IT teams, that affects policy enforcement and the ability to trace a response back to verified ground truth.

For operations teams, that affects response quality and user wait times.

What to check first

If a model starts pulling from different sources over time, check these in order:

1. Source history

   • Did the underlying raw source change, move, or get retired?

2. Index or compilation history

   • Was the compiled knowledge base rebuilt or re-ranked?

3. Prompt and routing

   • Did the system prompt, tool choice, or fallback logic change?

4. Model version

   • Did the model, provider, or orchestration layer update?

5. Permissions

   • Did a source become unavailable because of access rules or compliance filters?

6. Query wording

   • Did the prompt change enough to alter intent?

7. Ground truth

   • Does the answer still trace back to a verified source, or only to a similar one?

How to keep source behavior stable

If you need consistent answers, treat source control as a governance problem.

• Ingest raw sources into one governed, version-controlled compiled knowledge base.
• Keep source IDs, version dates, and retrieval timestamps tied to each answer.
• Score every response against verified ground truth.
• Review freshness and authority rules together, not in isolation.
• Route gaps to the right owners when the model cannot find a citation-accurate source.
• Monitor public AI answers over time if external representation matters.

That approach reduces drift because it makes the source path visible. It also makes it easier to explain why a model used one source last month and another source today.

Is source drift the same as hallucination?

No.

Hallucination means the model states something without support.

Source drift means the model may still be grounded, but it is grounding the answer in a different source than before.

That difference matters. A grounded answer can still create compliance risk if the source is outdated, weaker, or harder to audit.

Why this is a governance issue, not just a model issue

Most enterprises are deploying agents faster than they are governing the knowledge behind them. That gap is where misrepresentation, inconsistent answers, and liability show up.

A context layer fixes the part the model cannot manage on its own. It compiles the enterprise’s full knowledge surface into a governed knowledge base. It tracks citation accuracy against verified ground truth. It gives teams a way to prove what the agent used and why.

That is the control point.

FAQ

Why does a model cite different sources for the same question?

Because the retrieval layer, ranking rules, prompt, model version, or source availability changed. The question may be the same, but the source path is not.

Is it normal for source selection to change over time?

Yes. It is normal in systems with live content, fresh indexing, or changing access rules. It is not acceptable when the organization needs stable, citation-accurate answers and cannot explain the change.

How do you reduce source drift in enterprise agents?

Version-control the raw sources, compile them into one governed knowledge base, and score every response against verified ground truth. That keeps the model grounded and makes source changes visible.

What does source drift mean for AI Visibility?

It means your organization can be represented differently from one week to the next. If public AI answers change, your narrative control changes with them.

If you need stable, provable answers, do not treat the model as the source of truth. Treat the source layer as the system that needs governance. Senso is built for that layer.