What to ask your agent

Plain-language prompts work. These are the nine jobs people actually run, with the tools each one uses, so you can see what the agent is doing and tell when it has gone the wrong way.

Setting up

Onboard a project

Have the agent do the whole setup: brand, competitors, buyer prompts.

Run the Omnipresence MCP onboarding

Tools, in order: lookup and fetch for the onboarding process, project_context to see what is already set, project_context_update for the brand and voice fields, presence_manage_asset once per brand (yours with is_owned: true, then each competitor with its aliases), presence_manage_prompt with action: "create" per prompt, and memory_store with facet: "decision" for the choices it made.

What you get back: a configured project with prompts on a schedule, and a written record of why those competitors and those prompts. Ask "why those five competitors?" before you accept it.

Add prompts in bulk

You have a list of buyer questions and want them tracked.

Add these as weekly tracked prompts, topic "comparison", and make the last
one a control: [paste your list]

Tools: presence_manage_prompt with action: "create" per prompt, taking frequency (daily, every_3_days, weekly, biweekly or monthly), topic, samples, and is_control on the one you will never work. Then presence_tag with action: "add", entity_type: "prompt" to group them.

What you get back: the prompts created, with anything it refused and why. A duplicate is reused rather than added twice.

For a one-off study rather than ongoing tracking, ask for a research batch instead: presence_research_batch costs the whole set up front and never occupies a tracked slot. See research prompts.

Checking where you stand

The weekly visibility check

The standing Monday question, answered in one turn.

How did we do last week? Visibility Score with the band, the change on the
week, the per-platform split, and anything that moved outside the band.

Tools: presence_overview with from, to and mode: "grounded" for the headline, then presence_timeseries with metric: "mention_rate" for the shape, and presence_query with entity: "prompts" to find which prompts moved.

What you get back: the rate with its confidence band, the grounded against training split, how often search actually fired, share of voice, and the counts. Ask it to say explicitly whether a move is inside the band, because that is the difference between a result and noise.

Why is a competitor recommended and not me

The most useful question in the product, and the one that takes longest by hand.

For the prompts where we are not recommended, who is, and which sources are
they cited from that we are not on?

Tools: presence_query with entity: "prompts" to find the weak ones, presence_answers with a prompt_id to read what the models actually said, presence_query with entity: "citations" or entity: "sources" for the pages behind those answers, and presence_get_run for one run end to end.

What you get back: the competitors named instead of you, in the order they were named, and the pages the models drew on. Nearly always one or two sources that carry a rival and not you, which is the next job.

Doing the work

Find a source opportunity and claim it

Turning the source data into outreach, and recording the result.

What are the top unclaimed placement targets we are not on, by how often
models read them?

Tools: presence_query with entity: "opportunities" and only_unclaimed: true, or entity: "fanouts" with order: "opportunity" to rank by where the gap is worth closing rather than by raw frequency. Then presence_source_update with notes as you work it, and later with claimed: true and claimed_url once the placement is genuinely live.

What you get back: a ranked list with each source's role (directory, roundup, business homepage, user generated or regulatory), whether your brand is present, and how many times a model has read it. See source opportunities.

Audit a page and get the rewrites

Why a page is not being quoted, and what to change.

Score our pricing page against the fan-out queries it should answer, then
rewrite the weak passage and score the rewrite before I publish it.

Tools: presence_query with entity: "fanouts" for the queries the page should serve, fetch for the retrieval methodology, then chunkscore with the query and the candidate passages. The agent iterates on the draft until the passage clears.

What you get back: a score per passage against each query, and a rewrite that clears relevance 105 and n-gram 3 before it goes live. This is the loop that otherwise takes a week: write, publish, crawl, score, discover it does not answer the query, rewrite. See retrieval readiness.

Plan content from the methodology

Getting the method rather than an improvisation.

We are invisible on comparison questions. Find the right process for that
and plan the work against our actual data.

Tools: lookup with the problem in natural language, fetch with the ids it chose, project_context and project_files_list for the brand's own rules, and presence_query for the evidence. Then memory_store with facet: "topical_map" or "decision" to record the plan.

What you get back: a plan that follows a defined process and cites your own numbers, rather than generic advice. Ask which process it is following; an agent that cannot name one has not fetched it.

Log an optimization and check its impact

Closing the loop, which is what makes the next session useful.

We published the new comparison page today. Flag that query as being worked,
log it, and in two weeks tell me whether it moved.

Tools now: presence_set_optimizing with optimizing: true, which stamps the date every before-and-after is drawn from, then presence_log_optimization with title, detail, page_url and optimized_on.

Tools in two weeks: presence_query with entity: "fanouts", only_optimizing: true and include_sources: true for the work queue with its last logged change, plus presence_timeseries for the movement since that date.

What you get back: a dated record that survives the chat window, and a comparison with a defensible start line. Log work that has actually gone live, never work that is drafted or agreed.

Reporting

Export a report

Getting the data out, for a client or your own analysis.

Export last month's runs with the answers and what was cited, as xlsx.

Tools: presence_export with from, to, format (csv, xlsx or jsonl) and include (answer, retrieved, cited, fanouts). It returns a signed download link valid for 24 hours, with the file size and checksum; delivery: "inline" returns one page in the response instead, paged by cursor.

What you get back: one self-contained record per run, carrying the prompt, your own external_id, the engine, whether it searched, the fan-out queries, everything retrieved and what the answer actually cited, and the cost. Enough for a study or a client pack without opening runs one at a time.