eoma
Türkçe Log inCreate account

Docs › Concepts

What a run is

EOMA puts your questions to four AI engines on a schedule. Here is why the same question is asked over and over, and why one run never proves anything.

In 10 seconds

The short version

An AI engine is not a database. Ask ChatGPT "what are the best villa rental sites in Antalya?" three times and you can get three different lists. Nothing broke; that is how the technology works. It means a single answer tells you almost nothing, and a single answer treated as a measurement produces a number that swings for no reason and gets acted on anyway.

EOMA's answer to that is repetition. Your questions go out on a schedule, to four engines, week after week. Each individual answer is an anecdote. Several hundred of them, pooled over 28 days, is a measurement — and the same pool measured again next month is a trend.

That is the whole reason the product accumulates rather than reporting. In your first fortnight you have a snapshot. By week eight you have a direction, and the direction is the part that tells you whether the work you did was worth doing.

How it works

  1. A schedule fires. On every paid plan, the AI visibility collection runs once a week — four runs a month, on the same weekday each week. The spacing is deliberate: your monthly question allowance is spread across those four runs instead of being spent in one enormous tick, so a number always reflects several points in time rather than one afternoon.

  2. A quarter of your question set goes out, not all of it. Each run asks your monthly allowance divided by four — so a Starter run puts out about 8 questions, Growth 25, Scale 50 — and the remainder waits for the next run. Four runs, one month, your whole allowance asked exactly once. If you track fewer questions than your plan allows, the run is sized on what you actually track, never on the plan's ceiling.

  3. Each question is put to four engines. ChatGPT, Claude, Gemini and Perplexity, with the same wording, adjusted only where an engine will not accept a setting — Gemini takes no location fields, so the location has to sit in the question itself; Perplexity accepts a country but not a city. Each question is also asked in two modes, one where the engine may search the web and one where it answers from what it already knows. That distinction is frequently the finding: a brand the engines can look up but do not know is a different problem from one they cannot find at all.

  4. No question is asked twice until every question has been asked once. The rotation picks the questions that have gone longest without being asked, so a question you added last week is asked before one asked yesterday. That is a guarantee about coverage rather than a waiting period: it is what makes "four runs cover your whole set" true rather than approximately true, and it is why the rotation cannot get stuck on your five favourite questions.

  5. Answers are added, never replaced. A question asked in July and again in August produces two records under the same question, not one overwritten record. That is what makes a per-question history possible, and it is why editing a question's wording does not detach it from its own past.

  6. The scoring runs afterwards. Answers are read for who was named, which sources were cited, and whether the engine actually knew anything. Competitor names are folded together first — "Airbnb" and "AirBnB" are one rival — and only then is anything scored. Scoring before that step would count one competitor as two and quietly deflate everything.

What this means for you

Not every active question is asked every month, and that is not a fault. Your active set is a rotation pool, not a promise. If you add thirty questions mid-month, some of them will not be reached before the month turns over. Some may wait several months. There is no error, no warning and no health check for this, deliberately — a warning that fires constantly and correctly is one everybody learns to ignore.

Turning a question off and on again does not push it to the front of the queue. The 30-day clock survives deactivation. If you switched a question off five days after it ran, switching it back on does not buy it an extra send.

A small active set gets re-asked, not under-spent. If you narrow to five questions, those five are asked repeatedly. Your cost does not change; your data just gets repetitive.

One run is never evidence, and neither is one day's change. If your number moved and you changed nothing, the most likely explanation is that the pool moved: a run entered the 28-day window and an older one fell out. Compare weeks.

If nothing has run in 28 days, EOMA says so rather than showing you an old number. A blank score with "not measured" is the honest state. Data frozen weeks behind its schedule is worth reporting to us.

Common misunderstandings

People assume Actually
"Every question I have is asked every time." Each run takes a slice sized by your plan. The whole set is covered across a month of runs, and a question added late may wait.
"The answer changed, so the tool is unreliable." The engines genuinely answer differently between asks. That variance is the reason for repeated measurement, not evidence against it.
"A run is a report being generated." A run is EOMA spending real money asking real engines real questions. It produces new evidence, which is why it is paced rather than run on demand.
"I published a fix, so the next run should show it." The engines re-read the web on their own schedule. Weeks is the normal delay between a change going live and an answer changing.
"More runs would give me a better number." Beyond the current cadence, more runs mostly buy noise. The 28-day pool already holds several runs and well over a hundred answers for an active brand.
Reading this as a machine? The same page in Markdown: /docs/concepts/what-a-run-is.md

On this page