# Where did we lose ground in AI answers this week?

> For the complete documentation index, see [llms.txt](https://attensira.com/llms.txt).
> Every page on this site is also served as Markdown: append `.md` to any URL, or send `Accept: text/markdown`.

> Ask your assistant whether you lost ground in AI answers this week: the Attensira MCP call, the chain down to the prompts that moved, and what counts as real.

- Canonical page: https://attensira.com/mcp/use-cases/find-where-you-lost-ground-this-week
- Family: mcp-use-cases
- Last updated: 2026-10-05
- Author: Karl-Gustav Kallasmaa (https://x.com/kallasmaa)
- Claims as data: https://attensira.com/claims.json?slug=find-where-you-lost-ground-this-week

**In short:** Ask your assistant to call get_analytics with area=visibility and days=7 on the Attensira MCP server. It returns your mention and citation rates per engine with each change marked real or not. If nothing is real, stop. If something is, rank prompts worst first and read the answers behind the two or three that moved.

## What it showed on attensira.com (read 2026-10-04)

Asked in Claude: "Did we lose ground in AI answers this week?"

Call: `get_analytics area=visibility days=7`

No loss we can prove. Neither the workspace's change nor either engine's passed the significance test, so the chain stops here.

| Row | Value | Note |
| --- | --- | --- |
| All answers | 6.0% | n = 2,085 · change not proven |
| ChatGPT | 5.7% | n = 1,098 · change not proven |
| Perplexity | 6.3% | n = 987 · change not proven |

_The first call of the week, on our own workspace._

## Which call answers it?

```json
{ "tool": "get_analytics", "arguments": { "area": "visibility", "days": 7 } }
```

## What is the chain behind it?

1. **Read the workspace level first** Call get_analytics with area=visibility and days=7. This establishes whether anything moved at all. If every delta comes back real:false, stop here — there is nothing to explain, and explaining it anyway is how a team learns to act on noise.
2. **Rank the prompts worst first** Call get_analytics with area=prompt, days=7 and order=worst_first. Ranking by how badly you are doing surfaces the losses instead of averaging them away against the prompts you already win.
3. **Read the per-platform rows** The same visibility call already returns one row per platform. A drop concentrated on one answer surface is a different problem from a drop across all of them; get_analytics area=platform with a filter opens one platform end to end.
4. **Read the answers behind the worst movers** Call get_prompt on the two or three prompts that moved most. The returned answer text is the only place in the product where you find out why a model stopped naming you, rather than that it did.

## How do I read what comes back?

| What you see | What it means |
| --- | --- |
| Every delta comes back with real set to false | No change you can prove. It is not a decline and it is not a gain; the sample cannot separate it from noise. |
| Your rate flat while a competitor's rises | You did not lose ground — somebody else gained. That calls for a different response from a loss. |
| A drop on one model with the others flat | Most likely a change on that answer surface rather than a change in your content. |
| A rate that fell while n is in single digits | One or two readings moved it. Widen the window before treating it as a finding. |

## What can't this tell you?

- Nothing in this chain says why a model changed its answer. Providers swap the models behind an answer surface without notice, and a rewritten system prompt on their side looks identical to a content problem on yours.
- A seven-day window gives very few readings per prompt per model, and a rate built on a single-digit n swings for reasons that have nothing to do with you. Widen days before acting.

### What did it show on our own workspace?

No loss we can prove, which is the usual shape of a week. The right report was one line long, and the chain stopped after its first call. That is the point of starting at the workspace level: a null result costs one call, so a weekly ritual can afford to find nothing, instead of turning into a search for something to report.

### Why does a drop on one engine point away from your site?

A drop that shows on one answer surface and nowhere else usually means that surface changed, not your pages. Open it on its own with `area=platform` and a `filter`. The fix for a surface change is patience and a wider window; the fix for a page problem is a content change. The worst-first prompt ranking is what keeps a real loss on three prompts from being flattened by twenty that did not move.

### What does a mention rate not tell you?

How the answer described you. A mention that criticises you counts exactly like a recommendation, because mention rate measures whether you were named, not how. That is a separate reading, on the [Sentiment page](/features/sentiment).

When one prompt keeps showing up, it has [its own recipe](/mcp/use-cases/turn-a-losing-prompt-into-a-plan). If the movement was a rival's rather than yours, [check whether a competitor is gaining](/mcp/use-cases/check-whether-a-competitor-is-gaining). After running the chain twice by hand, [schedule it](/mcp/use-cases/schedule-a-weekly-visibility-digest).

## FAQ

### What does a delta of real:false mean?

It means the change is not distinguishable from sampling noise at the sample size collected. It is not a decline, it is not a gain, and it should not be reported as either. Attensira returns the flag precisely so that a weekly report can be written to ignore movement it cannot prove.

### Why does a rate come back as null instead of zero?

Null means the model was never queried in that window — the row is marked tracked:false — so there is no reading to average. Zero means it was queried and never mentioned you. Treating null as zero invents a measured absence that nobody measured.

### How short a window is too short?

Each prompt is read at most once per model per day, so a seven-day window holds at most seven readings per prompt per model. Each reading is 1 sample on Starter and 3 on Growth and up, so that is at most 7 answers, or 21. A rate on that few answers moves several points when one answer changes.

### Do I need write access for this recipe?

No. get_analytics and get_prompt are both read-only tools, so a read-only API key runs the whole chain, as does any signed-in connection. Nothing here spends credits.

### Can I run this on a schedule instead of by hand?

Yes — that is what create_automation is for, and the weekly digest recipe covers it. Write the instruction to name the areas and windows explicitly and to ignore deltas where real is false.

## Sources

Each entry below is one claim on this page. `retrieved` is the date the exporter or a human last saw the source say this, and nothing more: it is not a publication date and not a promise of re-verification. `cite` is the anchor on the canonical page where this claim is rendered.

### sampling-cadence

Attensira's MCP recipes state that sampling runs at most once per prompt per model per day, so a plan executed today shows its earliest possible signal tomorrow.

- source: https://docs.attensira.com/mcp/recipes (docs.attensira.com)
- retrieved: 2026-10-05
- cite: https://attensira.com/mcp/use-cases/find-where-you-lost-ground-this-week#claim-sampling-cadence

> Sampling is at most once per prompt per model per day, so a plan executed today shows its earliest possible signal tomorrow and a trustworthy one much later.