Find out which questions name you and which name a competitor. Publish the page that answers the ones you are losing, with every fact checked by a person. Ask the same questions again and see what moved. Each stage is below in plain words. Flip nerdy details for exactly how it is done.
Before anything is measured, we get the facts straight. Where you are, what you do, who works there, what you take. You confirm the list. From then on, every AI answer about you is judged against it.
Share of voice corrected from 18.4 to 11.9 percent with the mention rate unchanged. That is how you know the change was surgical.
Everything downstream is judged against what we collect here, so the first stage is the least glamorous and the most important. It produces three things: a fact sheet, a market definition, and a competitor set. All three are yours, all three are visible on your portal, and nothing gets measured until you have looked at them.
The fact sheet is the list of true statements about your business that an AI engine could get wrong: services you offer and ones you have stopped offering, every location with its address, the names of the people who work there, insurance panels and payment types, hours, credentials, the year you opened. We draft it from your site and your Google Business Profile. You correct it. From then on it is the standard every claim is judged against, and nothing you have not confirmed is ever treated as fact, including by us.
The market definition matters more than it sounds. The same question asked from two cities returns two different answers, and a place name is often ambiguous. A San Marcos clinic measured without a state is partly measured against San Marcos, Texas. We found that in our own data, on a real client, four questions deep, and now every panel carries its market in the request itself rather than in the wording of the question.
The competitor set starts with the names you give us and then gets checked against reality. Before any competitor is tracked, we confirm it actually appears in answers about your buyer questions. A rival you worry about who never shows up in an AI answer is not a competitor in this channel. A regional health system you never think of as a rival, that appears in one answer in five, is. In one calibration the set went from ten names to twenty-three, and share of voice corrected from 18.4 to 11.9 percent without the mention rate moving at all, which is how you know the change was surgical.
One rule from this stage carries through all of them: a founder’s name is not the organization’s name. An answer that names the person who invented your field is not an answer that names your practice. The two are measured separately and reported side by side, because we once counted them together and it inflated a client for months.
Your next customer types a question into ChatGPT before they know your name. We write down twelve to twenty of those questions, in their words. You read every one and add what we missed.
The panel mixes shapes on purpose, so the map shows the whole shape of demand and not the flattering slice.
A measurement is only as good as its questions. The panel is the fixed set of buyer questions, usually twelve to twenty, that we ask every engine every month. Not your name. The problem. "Best physical therapist for runners in Encinitas." "Trauma therapist near me who takes my insurance." "Hospice care options in North County." The questions a person types when they do not know you exist yet.
We draft them from your market, your services and your competitors, then show you every one and ask what is missing. The questions you never thought to ask are usually the point. A practice that leads with athletic performance and has never asked whether the engines name it for sports rehab is about to learn something.
Each question is written in the shape buyers actually use, which differs by engine and by intent. A best-of question, an alternatives question, a does-this-place-do-X question and a plain informational question pull different kinds of sources and produce different kinds of answers. The panel mixes them on purpose so the battle map shows where you hold and where you do not across the whole shape of demand, not just the flattering slice.
Once you approve it, the panel is frozen and given a version number. Question text is the measurement; change the wording and every trend line restarts. So a panel changes only at a month boundary, with the change logged, and the old version stays readable. When we fixed the San Marcos ambiguity, the corrected panel shipped as a new version, dated, with the four affected questions named, so nobody reading the ledger later mistakes a wording fix for a market move.
You can propose new questions any time. They join the next version. Five of them can also be put on the Monday watch, where they are re-asked weekly between the monthly runs, which is the fastest honest signal we know how to give.
Every month, all eight engines get every question, more than once. Every answer is saved word for word. Nothing on your portal is a score without the actual answer behind it.
Excluding answers where no business was named moves every client up between one and eight points. The methodology card prints which denominator you are looking at.
Every question on the panel goes to eight engines: ChatGPT, Perplexity, Claude, Gemini, Grok, Google AI Overviews, Google AI Mode and Copilot. Each engine gets each question more than once in a run, because an engine asked the same thing twice will sometimes name you once. The verbatim answer is kept every time. That saved answer is the receipt, and everything else on the portal points back to one.
The engines that can search, search. ChatGPT and Gemini are called with web search on, so their answers arrive with the sources they actually pulled rather than from training memory. That distinction is not academic. Before we turned it on, those two engines were mostly declining to name anyone for local questions, and we were recording those refusals as losses. Two of eight engines were measuring a refusal rate, and every client in the book looked worse than they were. The fix was an engine version change, dated in the change log, and the numbers before and after it are not compared to each other.
Answers that name nobody are classified separately. Sometimes the engine declines to rank anyone, sometimes the question was informational, sometimes Google simply does not render an AI Overview. Nobody wins those answers, including your competitors, so they are reported as no-names or no-answer rather than as you losing. The honest denominator moves every client up between one and eight points, and it is printed on the methodology card so you can see which denominator you are looking at.
A citation is a page the engine read. A search result page is not a citation and is never counted as one. Redirect links are resolved to the real hostname, because a source map full of Google redirects is useless for outreach. The rule that decides what counts is versioned, and its change log is printed on the same card.
The measurement runs once a month on the whole panel, and on demand for a single page’s questions when you ship something and want to know that day whether it moved. The monthly run is the trend. The on-demand run is a reading. Both are labeled as what they are.
One line appears under every number we publish, and it belongs here too: API responses approximate but do not exactly match consumer app answers; we report these numbers as trend and share of voice comparisons, never as absolute truth.
Northshore Physical Therapy in Encinitas runs a post-surgical ACL program with return-to-sport testing. Coastal Motion PT and Encinitas Sports Rehab also offer rehab...
For runners, Coastal Motion PT is frequently recommended for gait analysis and return-to-running plans. Movement Lab is another option in Carlsbad...
In California you can see a physical therapist without a referral for a limited period. Look for a clinic that treats foot and ankle conditions...
One screen that answers the question you actually have. Your questions down the side, the eight engines across the top. Green where an engine names you. Red where it names a competitor. Grey where it names no one.
Northshore Physical Therapy, fictional sample. The seven nobody holds are the cheapest ground on the map.
The battle map is the measurement laid out so an owner can read it in a minute. Rows are your questions. Columns are the engines. Each cell is held (an engine named you), lost (an engine named a competitor and not you), or nobody’s (no business was named). The receipt is one click from every cell.
Across the top sits coverage: the number of questions where at least one engine names you, out of the panel. Under it, who holds the rest. In the sample on our home page, Northshore Physical Therapy holds eighteen of forty-six questions, a competitor holds twenty-one, and nobody holds seven. Those seven are the cheapest ground on the map, and the twenty-one tell you exactly who is getting the call today.
Three things are scored separately because they are different things: a mention, a citation, and a recommendation. Being named in passing in a list of six is not the same as being the answer, and an engine linking to your page is not the same as an engine telling someone to call you. The recommended-first rate is the number most owners end up watching.
The map also shows how sure it is. A cell where the samples disagreed shows the count on the tile, so a result won three times out of three is never displayed as if it were the same as one won once. Any cell that split, or that changed hands since last month, is asked again until it is at least five samples deep. A move counts as confirmed only when the answer held across that deeper set. Anything shallower is drawn provisional, in a dashed edge, and is never counted as proof. In September that re-sampling retired about a quarter of the flips a three-sample read had claimed. We would rather show you fewer wins that are real.
Every battle map carries a methodology card: the panel version, the market it was measured from, which engines accept a location and which answer nationally, the sampling depth, the scoring rule, and the date. It is the same card whether you are on The map or Done for you. The map is the product, and the card is what makes it a measurement instead of a chart.
Being named is half the job. Being named correctly is the other half. We ask the engines about you by name and check every claim against your fact sheet. Wrong hours, a service you dropped, an insurance panel you left years ago. Each one gets traced to the page that taught it, when there is one to find.
All 48 got a documented fix plan.
Half the risk is being mentioned wrong. Monitoring tools count mentions and score their mood. We read them. Alongside the buyer questions, every run asks each engine about you by name, and every factual statement in those answers becomes a claim: a service, a location, a person, an insurance panel, a price, an hour, a year.
Each claim is judged against the fact sheet you confirmed in intake. Correct, incorrect, or unverifiable. Nothing is judged from memory, and nothing you do not publish yourself is ever published by us in a correction. In one month across the book, 398 claims were checked and 48 came back flatly wrong. Some were small. Some described a real estate brokerage as a school bus company, or told parents that a childcare center was a hospital donation page. All 48 got a documented fix plan.
A wrong claim almost always has a page behind it. Engines cite what they read, and what they read is often a directory listing from 2022, a press mention from a previous owner, or a page on your own site that was true four years ago. Tracing the claim to its source is the diagnosis. Correcting the source is the fix. You cannot edit an engine, but you can change what it finds.
Sentiment is a different question from accuracy and it is the one every other tool answers. The trouble is that a false statement delivered warmly scores as positive. An engine can confidently tell a patient that you do not take Medicare, in a friendly tone, and a sentiment tracker will file it under good news. A wrong claim is not a mood. It is a customer who never calls to check.
On Done for you a weekly watchdog runs the branded questions between the monthly audits, so a new wrong claim is caught in days rather than weeks. The accuracy tab on your portal lists every claim, its verdict, its source, its fix and its re-measure date. The public version of that, anonymized, is the wrong-answers library on this site.
For every question you are losing, three answers: who has it, why they have it, and what page would have to exist for an engine to name you instead. That third answer becomes the prompt for the page.
Same underlying data as a citation dashboard, except one of us hands back a to-do list.
A diagnosis you cannot act on is a chart. So every lost question on the map gets the same three answers: who holds it, which sources put them there, and what would have to exist for an engine to name you instead. The third answer becomes a brief.
The source map is where the diagnosis starts. For each lost question we look at what the engines cited when they named your competitor. Your own domain and every competitor domain are excluded by rule, because neither is an outreach target. What is left is scored by influence, opportunity and attainability. A directory you are not listed in, an association page that names three members and not you, a local publication that quoted your rival. Same underlying data as a citation dashboard, except one of us hands back a to-do list.
Sources are reported by question shape, because engines behave differently by shape. A best-of question pulls forums, review sites and video. An alternatives question pulls comparison pages. A does-this-place-do-X question pulls your own site and your Google Business Profile. The kind of page that wins each one is different, and the brief says which kind you are writing.
The brief itself is tied to one lost question. It states the question, the engines that lost it, the competitor that holds it, the sources they hold it with, the facts from your sheet that the page must contain, the facts it must never claim, and the shape the page should take so an answer engine can lift from it: a direct answer up front, every target question covered, your name as the entity, no figure outside your verified facts. It is written by a person and it reads like one wrote it. On Done for you every lost question gets one. Teams with their own writer can have the briefs alone, ready to hand to whoever writes for you.
Wrong claims get a brief of their own, a correction asset: the true statement, the source it needs to reach, and the request to send. Most corrections are one email or one form. The hard part was knowing which one.
The page gets written, checked against your facts, and published. On Done for you, we write it and a person reads it before it ships. You approve every word. On The map, you copy the prompt and write it yourself.
Both scores are advisory. A person still decides.
This is the stage where the plans actually differ. On The map you take the diagnosis and write from it yourself. On Done for you we brief it, write it and place it, and you approve every word before anything is published. Teams with a writer can ask for the middle path, we prep and you cook: the brief and the draft check, your writer does the rest. The measurement, the map, the audit and the ledger are identical either way. What changes is whose hands are on the keyboard.
Before a draft ships, whoever wrote it can paste it on the Briefs tab and get two scores. The first is how human it reads, scored against the published signs of machine writing, by density. The second is whether an answer engine can lift it: a direct answer up front, every target question covered, your name as the entity, no figure outside your verified facts, no never-true claim. Both scores are advisory. A person still decides.
Nothing on Done for you publishes without your approval in the portal, and every unverified fact in a draft stays a visible placeholder until you resolve it. We would rather ship a page a week late than ship a page that teaches an engine a new wrong claim about you. The accuracy layer exists because other people’s pages did exactly that.
Corrections go to the source the engine trusts, which is rarely your own site. A directory update, an association profile, a Google Business Profile edit, a note to a publication that got a fact wrong in 2023. Each one is logged with the date it was sent and the date it was confirmed, because the re-measure has to know when to expect movement.
When a page ships, you mark it shipped and the date lands on your trend chart. Spend a credit and that page’s questions are asked again the same day on all eight engines. That reading does not change the month’s trend. It tells you whether the engines have seen the page yet, which is a different and useful thing to know.
Engines decide what they cite. Nobody can promise placement, and anyone who does is guessing. What we promise is that the page will exist, it will be correct, it will be shaped so an engine can use it, and the next measurement will say whether it did.
Next month, same questions, same engines. The ledger shows what moved after the work went live, with the saved answers behind it. A change is evidence to read, not proof that one page caused it. One good week is noise, so a win only counts once it holds.
Single-month moves are noise. Flips take 60 to 90 days after indexing, and an engine can name you on a Tuesday and not on a Thursday. Trend, never truth.
A month later the same panel goes to the same engines, and the ledger records what moved. Every line names the question, the engine, what it was, what it is now, and what shipped in between. Not a score that went up. A question you did not hold on Perplexity in August that you hold in September, with the page that was published on the 14th sitting next to it.
Single-month moves are noise. Flips take 60 to 90 days after indexing, and an engine can name you on a Tuesday and not on a Thursday. So a flip is confirmed only when it holds across the deeper sample described in the battle map stage, and the ledger marks the difference between confirmed and provisional in plain sight. Trend, never truth. We would rather show you a smaller number that stays true than a bigger one that evaporates.
The movement report is the ledger with a narrative: what shipped, what moved, what did not, and what next month’s briefs are aimed at. It goes out monthly on both plans. On Done for you it also carries the outcomes view, which joins Google Analytics to the visibility numbers so AI referral sessions and conversions sit next to the map. That number is labeled a floor, not a total, because engines send people who arrive typing your name into a browser and never show up as a referral.
The ledger is also how the process corrects itself. When two independent instruments disagreed about one of our clients, the ledger is where we found that a founder’s name had been counted as the organization for months. The fix, the restated numbers and the date all went into the ledger, because a measurement that cannot record its own mistakes is not one you should trust with anyone else’s.
Then the loop turns again. Intake facts get updated when something changes. The panel picks up the questions you proposed. The competitor set gets checked against the new answers. That is the whole product: not a report, a loop, with a ledger that gets longer every month you run it. Every stage of it is on this page because a prospect who reads it understands why the monthly work is worth paying for, and a client who would rather run it themselves can.
Yes. Every stage is on this page and on the portal. The map gives you the measurement, the battle map, the accuracy audit and the ledger, and you do the writing and the corrections from the diagnosis. Most owners look at the eight stages and choose done for you, but the choice is yours.
Eight: ChatGPT, Perplexity, Claude, Gemini, Grok, Google AI Overviews, Google AI Mode and Copilot. Every question on the panel goes to all eight, several times, every month.
Flips typically take 60 to 90 days after the engines index a change. Single-month moves are treated as noise until they hold across a deeper sample. The ledger marks confirmed and provisional movement separately.
No. Engines decide what they cite. Eidomai guarantees that every answer behind the measurement is saved and shown to you, corrections go to every source we can identify, the pages are correct and shaped so an engine can use them, and the next measurement reports what happened.
The verbatim answer an engine gave, saved at the time it was given, with the sources it cited. Every number on the portal points back to one.
API responses approximate but do not exactly match consumer app answers; we report these numbers as trend and share of voice comparisons, never as absolute truth. Who does what on each plan: Who cooks.
Free check on ChatGPT and Claude. Confirm your email, get a link to your results; no call required. Want the full eight-engine walkthrough? Book a demo and it is ready before the call.
It starts with an eye. It ends with AI. Eidomai, by Nimbletoad.