Position drift is the gradual or sudden shift in where your brand appears within AI-generated responses over time: for example, moving from being the first-named recommendation to appearing third in a bulleted list, or dropping from a prominent endorsement to a brief footnote mention. Detecting position drift early is critical because small ranking shifts compound into significant visibility losses.
Why position drift happens
AI responses aren’t static. They change when:
- Model updates: AI providers retrain or fine-tune their models, which can shift which brands they associate with a category
- Competitor gains: a competitor earns new citations, press coverage, or structured data that increases their authority relative to yours
- Content decay: your content becomes less fresh or loses backlinks relative to the competition
- Query reformulation: AI engines change how they interpret a query type, favoring different content signals
- Engine behavior changes: providers adjust retrieval, ranking, or citation policies
Most position drift is gradual and goes unnoticed without systematic monitoring.
Position tiers in AI responses
These are the five tiers LLM Metrix classifies every answer into, by where your brand sits among all the brands that answer names, not just among the competitors you track.
| Position | Description | Relative value |
|---|---|---|
| First | Named ahead of every other brand in the answer | Highest |
| Prominent | Second or third (in the recommendation, not the long tail of it) | High |
| Mid | Fourth through seventh, typically mid-list | Medium |
| Fine-print | Eighth or later, or named with no discernible order at all | Low |
| Absent | Brand not named in the response | None |
A single-tier slip (prominent → mid) can represent a 30–50% drop in effective brand impression from that query.
Detecting position drift in LLM Metrix
LLM Metrix re-runs your tracked prompts on your plan’s monitoring cadence (daily on paid plans, weekly on free) and classifies where your brand landed in each answer using the tiers above. See pricing for the cadence on each plan.
Position itself is history you read rather than a notification that fires:
- Position trend charts on Rankings: two plots over your scan history. One is your average position across the answers that named you; the other is the share of delivered answers sitting in each tier, stacked. The share plot is the one to trust when your prompt set has changed, because raw tier counts rise whenever you add prompts and that reads as an improvement it isn’t.
- A per-scan position breakdown: for the most recent scan, the tier for every prompt and engine, so you can see which prompts are pulling the average around.
- Score and share-of-voice trends: the same history for the composite visibility score and for your share against the competitors you track.
Two things worth knowing so you don’t wait for something that isn’t coming.
Position drift IS alerted: three forms, one switch. The position-change alert compares your average answer position between scans and fires when it moves by your threshold (default: a position or more). When the average-rank branch cannot speak (no previous scan, no ranks, or a move under the threshold), a fallback form compares your first-mention share and fires when it drops by your threshold (default: 10 points). The third form is the per-query one: when an individual (engine, prompt) row crosses a position band between scans (first → prominent, mid-list → fine-print, named → absent), it is named in a position-class-transition alert. All three live behind the same “Position change” switch in Settings → Notifications. The fallback forms exist because “you are losing the top spot” is the commercial reading of drift: an average can hold steady while individual queries slip a band, and a per-row slide says something the project-level numbers miss.
Position does reach one alert indirectly, and it is worth knowing precisely because it can be misread. Prominence is one of the three inputs to the visibility score (being named first is worth full credit, second or third rather less, eighth or later very little), so a widespread slip drags the score down and can trip the score-change alert. That email will say your score fell. It will not say position was the cause, and it says nothing at all about a slip small enough or narrow enough to leave the score inside its threshold. So: don’t read a quiet inbox as “our position held”. Read it as “the score, share of voice, citations and sentiment did not move by more than their margins”, which is a different and much weaker statement. A slide that stays inside one band (second to third place, both prominent) is still something you go and look at.
The Answer Archive diffs answers for you. Verbatim answer text is stored for every prompt, engine and scan, and the Answer Archive on Rankings puts each answer next to the prior scan’s (additions and removals highlighted, claims included), so last week’s wording sits beside this week’s without any manual comparison. You can also read raw answers in each scan’s detail or pull your scan history from the public REST API.
So the workflow for drift specifically is a review rather than a reaction: watch the position trend for a drop, then open the Answer Archive diff for the prompts that moved and read what changed.
What to do when you detect position drift
- Identify the affected prompts: is the drop isolated to a few prompts or spread across your tracking set? The per-scan breakdown gives you the tier for each prompt and engine.
- Check competitor gains: has a specific competitor recently published content, earned press, or changed their positioning in that topic area?
- Audit the cited sources: what sources is the AI citing ahead of yours in the scans where you slipped? Those are the pages you need to match or outperform.
- Prioritize high-volume prompts: a drop on a prompt with real search demand behind it deserves faster remediation than one on a rarely-asked question.
- Work the recommendations: LLM Metrix generates GEO recommendations from each scan, grounded in the prompts and engines where you were missed, the competitors named instead of you, and your entity presence in the knowledge graphs. They target the reasons an engine reached for someone else; they are not a position-recovery plan computed from your rank history, which the product does not have.