MCP recipes

Each recipe below is a real question someone asks on a Monday morning. Every one gives you the single call that mostly answers it, the longer chain when the single call is not enough, a table mapping what you see to what it means, and — the part that matters — what the numbers genuinely cannot tell you. All calls assume you are connected to https://mcp.attensira.com/mcp with a workspace API key. See the tool reference for parameters.

Find out where you lost ground this week

Start with workspace-level rates over a short window, then drill into the prompts that moved.

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

The chain

get_analytics area=visibility days=7

Establishes whether anything moved at all at the workspace level. If every delta comes back real: false, stop — there is nothing to explain.

get_analytics area=prompt days=7 order=worst_first limit=20

Ranks prompts by how badly you are doing on them, so the losses surface first rather than being averaged away.

get_analytics area=engine days=7

Splits the same window by model. A drop concentrated on one model is a different problem from a drop across all of them.

get_prompt on the two or three worst movers

Returns the actual answer text, which is the only place you find out why the model stopped naming you.

How to read it

What you see	What it means
Delta `{real: false, value: null}`	No change you can prove. Not a decline. Do not report it as one.
Your mention rate flat, a competitor’s rising	You did not lose ground — someone else gained. Different response.
Drop on one model, others flat	Likely an answer-surface change on that model, not a content problem.
`value: null` on a model row with `tracked: false`	That model was never queried. It is not a 0%.
Rate down but `n` is 7 or fewer	One or two bad readings can move this. Widen `days` before acting.

What this cannot tell you. A seven-day window on Starter sampling gives you very few readings per prompt, and a rate built on single-digit n swings hard for reasons that have nothing to do with you. Nothing here tells you why a model changed its answer: model providers swap the underlying models behind a surface without notice, and a rewritten system prompt on their side looks identical to a content problem on yours. You are seeing what changed, not who changed it.

See which sources the models trust in your category

Source share tells you which domains the models actually pull from when answering your prompts.

{ "tool": "get_analytics", "arguments": { "area": "source", "days": 90, "order": "best_first", "limit": 50 } }

The chain

get_analytics area=source days=90 best_first

A long window, because source share is a distribution and short windows make it lumpy. best_first puts the most-cited domains at the top.

get_analytics area=source days=90 filter=<one model>

Repeat per model. Perplexity and ChatGPT lean on visibly different domain mixes, and a category-wide average hides that.

get_analytics area=page days=90

Shows where your own domain lands in the same window, so you can see whether you are in the mix at all.

How to read it

What you see	What it means
One domain above roughly a fifth of all citations	Models treat it as the reference for your category. Being absent from it costs you.
A long flat tail of many small domains	No consensus source. Coverage breadth matters more than any single placement.
High source share, low `AvgPosition` index	Cited often and cited early in the list. The strongest signal available here.
Your domain absent entirely	You were not cited in this window. Not proof you are uncitable.

What this cannot tell you. Source share is citations to a domain over all citations in the window — it is not traffic, authority, or influence on the answer text. A model can name your brand in prose while citing a competitor’s blog, and a model can cite a page it disagreed with. The list also reflects your prompts, not the category: if your tracked prompts skew to one topic, the source mix skews with them. Widen the prompt set before calling it a picture of the category.

Check whether a competitor is gaining

{ "tool": "get_analytics", "arguments": { "area": "competitor", "days": 30 } }

The chain

get_analytics area=competitor days=30

Gives each tracked competitor’s mention rate on the same denominator as yours.

get_analytics area=visibility days=30

Your own rate over the same window, so you can compare movement rather than levels.

get_analytics area=prompt days=30 order=worst_first

Finds the specific prompts where the gap lives. Competitors rarely gain evenly.

get_prompt on those prompts

Read the answers. What the model says about them is the actual competitive intelligence.

How to read it

What you see	What it means
Their rate up, yours flat, both deltas `real: true`	They gained. You did not lose. The market grew around you.
Both rates up	The models are naming more brands per answer, not reshuffling ranks.
Their rate up, yours down	The closest thing to genuine displacement — and worth reading answers for.
Rates summing well past 100% across brands	Expected. Answers name several brands, so these never sum to 1.

What this cannot tell you. These are not shares of a fixed pie. Competitor mention rate uses the same denominator as yours, so it measures how often a brand appears in your prompts’ answers — nothing about their overall market presence, spend, or growth. A competitor you have not registered with add_competitor is invisible here and will read as absent, which is not the same as not appearing. There is no sentiment measurement, so a mention that trashes them counts exactly like a recommendation.

Turn a losing prompt into a plan

The one prompt you keep losing is the most actionable object in the product.

{ "tool": "get_prompt", "arguments": { "prompt": "best project management tool for agencies", "days": 90 } }

The chain

get_prompt with the exact text or id, days=90

Returns full answer text — not the 400-character snippets search returns — plus per-model rates.

get_analytics area=source filter=<this prompt or its topic>

Shows which domains the models leaned on when answering it. That is your target list.

get_page on any of your URLs that did get cited

Tells you what already works, which is a better template than guessing.

ask, passing what you found

Costs credits, and earns them here: the agent reasons across the slices you just pulled and drafts a plan. Poll get_session if it returns status: "running".

How to read it

What you see	What it means
Competitors named, you absent, citations to third-party lists	A placement problem. You need to be on those lists.
You named but never cited	The model knows you and does not reach for your pages. A content problem.
Cited but with a high `AvgPosition` index	You appear late in citation lists. Depth of coverage, not existence, is the gap.
Mention rate 0 with `n` above 20	A real, measured absence worth working on.

What this cannot tell you. Nothing here predicts that a change will work. There is no attribution path from a page edit to a later mention — you can only change something and re-measure after enough calendar days have passed to build a new n. 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. Answers also vary run to run at the model’s discretion; a single collected answer is an anecdote.

Set up a weekly report

{ "tool": "create_automation", "arguments": {
  "name": "Monday visibility digest",
  "instructions": "Summarise the last 7 days: workspace mention and citation rates with deltas, the five worst prompts, per-model split, and any competitor whose rate moved with real:true. Ignore deltas where real is false.",
  "schedule": "weekly on Monday"
} }

The chain

list_automations

Check what already exists. Two automations doing the same job bill twice.

create_automation

Free to create. Write the instructions to name the exact areas and windows you want, and to say explicitly that unproven deltas should be ignored.

run_automation with an idempotency_key

Test it once now rather than waiting a week. The key prevents a retry from producing a second billed run.

get_automation

Read the run history to confirm it produced what you expected before trusting it unattended.

How to read it

What you see	What it means
Digest reports every delta as a change	Your instructions did not exclude `real: false`. Rewrite them.
A week with no findings	Normal. Most weeks contain no statistically real movement.
Numbers differing from the dashboard	Check the window: 7 days against a 30-day default is a different slice, not a bug.
A run missing from the history	The scheduled run did not execute. Re-run manually with a fresh key.

What this cannot tell you. A weekly cadence over a daily sampling floor means each report rests on at most seven readings per prompt per model — three times that on Growth and Business, since sampling depth is three there. That is thin for week-over-week comparison, and a digest that dutifully reports movement every week is reporting noise. Credits and the response budget are currently inconsistent in the product, so do not assume a schedule that fits this month’s balance fits next month’s.

Audit which pages actually get cited

{ "tool": "get_analytics", "arguments": { "area": "page", "days": 90, "order": "best_first", "limit": 100 } }

The chain

get_analytics area=page days=90 best_first limit=100

Ranks your own URLs by citation activity. Long window, because page-level citations are sparse.

get_page on the top few

Shows which prompts and which models each page earns citations from.

get_analytics area=bot days=90

Shows AI crawler activity against your site — which pages were fetched at all.

How to read it

What you see	What it means
Page cited often across many prompts	A hub. Keep it current; it carries disproportionate weight.
Page crawled but never cited	Reachable and not chosen. A content problem, not an access one.
Page never crawled and never cited	Start with access — nothing can cite a page it has not fetched.
Citations concentrated on one model	That surface favours your format. Others may need different structure.

What this cannot tell you. Crawlers are not models: a crawler fetches pages, a model answers prompts, and a fetch is not a citation or a plan to make one. Crawler data also has a hard technical limit — AI crawlers do not execute JavaScript, so a client-side beacon, however written, has never seen one and never will; by the time the script runs, the only visitors left are humans with browsers. Use the install instructions the product generates at Settings for server-side collection. Finally, AvgPosition is a mean index within a citation list per source domain — it is not a rank against competitors and cannot be read as one.

Where to go next

Before acting on any of these numbers, read Reading the numbers for how null, zero, small n, and unproven deltas differ. Full parameter detail for every tool is in the tool reference.

Get started

What you track

Reading the numbers

Pages and AI traffic

The agent

MCP

Account

Reference

Find out where you lost ground this week

See which sources the models trust in your category

Check whether a competitor is gaining

Turn a losing prompt into a plan

Set up a weekly report

Audit which pages actually get cited

Where to go next

​Find out where you lost ground this week

​See which sources the models trust in your category

​Check whether a competitor is gaining

​Turn a losing prompt into a plan

​Set up a weekly report

​Audit which pages actually get cited

​Where to go next

Find out where you lost ground this week

See which sources the models trust in your category

Check whether a competitor is gaining

Turn a losing prompt into a plan

Set up a weekly report

Audit which pages actually get cited

Where to go next