Methodology

This page explains how NameIO counts what AI answers say about your store: what each figure measures, where the answers come from, how a question gets its funnel step, how sure a figure is, and when NameIO calls a change. The figures, tables and rules below are printed from the code that counts your answers, so they always match it. A dated changelog at the end lists every change to the method.

What we measure

NameIO asks the questions shoppers ask AI and reads every answer. Each question belongs to one funnel step: Awareness, Consideration or Conversion. For each step, on each engine and on all engines together, NameIO counts three things:

Named and share of brand mentions come with a likely range, or as counts when there are few answers. All three only use questions that don't name your store. These figures describe what AI answers say. They are not visits or sales.

What one answer is

An answer is what one engine said to one question in one check. NameIO keeps its text, finds every brand it names and reads the words around each name. A question asked again gives a new answer, and it can differ.

How brands are found in an answer

No AI model reads your answers for NameIO. Word rules find the brands, so the same answer and the same brand list always give the same brands. Your store and the competitors you track are matched by name. Other brands are found from how the answer is laid out, such as list items, bold names, headings and tables, and are counted once they pass the checks below.

Brand finder version 4, on a labelled set of 77 answers with made-up brands, written the way each engine writes, last run on Nov 11, 2026:

How the answer names brandsAnswersPrecisionRecallBar to pass
Lists, bold names and headings5798.0%93.4%Precision 95%, recall 85%
A table only6100.0%80.0%Reported, not gated
Sentences only14100.0%88.5%Reported, not gated

Precision is how many of the names found are brands a reader would count. Recall is how many of those brands were found. Retailers count here as brands, in the names found and in the labels, though shares leave them out.

On the structured answers, recall is 30.0% for brands named under one question only, and 87.7% when each question is read only with the brands found under the other questions.

A name counts as a brand once you mark it as one, once it links to a site with its name and is also written outside the link, once an answer leads with it and names it again, or once it leads under 2 different questions with at least one of those leads a list item, a table row or a link. Unless you mark it, a name that any answer used as a common word never counts. Review sites and publications never count; retailers are sorted onto their own line. A name written only as a link to a site with its name, or found only in a link's address, a citation or an image, is not counted.

Names are sorted by word rules, not by an AI model. A product written right after its brand counts as that brand; specs, section labels and joined phrases are not names; a site that sells other brands' products is a retailer. Names the rules can't sort count as competitors until you sort them. Your choices always win.

You can correct the list at any time: mark a name as Not a brand, or as Same as another brand.

Where answers come from

NameIO asks your questions on each engine you track and saves the answers. A data service that NameIO pays collects most of them: ChatGPT and Gemini answers are captured from their apps, as a shopper sees them; Google AI Overviews come from Google search results; other engines' answers come from the engine's own service for other apps (an API), so they can differ from what its app shows. Where answers come from has the details for each engine.

The three steps and the rule

Shoppers ask AI at every step of buying, so each question is sorted into one step:

Each step is its own set of questions, not the same shoppers moving through. Word rules sort the questions, with no AI model:

These are stage rules version 1. The rules run in this order, and the first one that fits decides:

RuleStepFires onConfidence
conv.trustConversionA named brand or store with a trust word: safe to order from, safe to buy from, safe, legit, legitimate, scam and 16 moreHigh
conv.purchaseConversionWhere to buy or order (where to buy, where can i buy, where to order, where can i order, where can i get, where to get and 8 more); about a named brand, discount, coupon, promo, promo code, voucher, deal and 18 more; about one named brand, its price or whether it is worth it (price, cost, how much, worth it, worth the money, worth the price and 2 more); with no brand named, a deal on a product type (deal on, discount on, coupon for)High about a named brand; low with no brand named
cons.compareConsiderationvs, versus, difference between, compare, alternative, instead of and 6 more; over about a named brand; two brands named with or, and, better, whichHigh
cons.shortlistConsiderationrecommend, review, top rated, rated, rating, trusted and 10 more; with a product type, best, top, which, who makes, who sells, looking for and 5 more; about a named brand, quality, any good, good; with no brand named, buy, purchase; a price used as a filter, such as under $50High
aware.learnAwarenessNo brand named, with how, why, what causes, what is, what are, what makes and 25 more; or a problem stated without a product type, such as help with, get rid of, deal with, struggle with, suffer from, i have and 5 moreHigh
fallbackAwareness with no brand named; Consideration with a brand namedAny question no rule above fitsLow: the app marks the step with a dot

Where two rules could fit:

Setup questions keep the step they were written for. For a store that sells one product type, they start as 4 Awareness, 11 Consideration and 5 Conversion questions. A step you set yourself stays when the rules change or the question is reworded.

You can change a question's step on the Questions tab. Your earlier weeks are then counted again with the new step, free.

The formulas

What each figure estimates, and what it leaves out:

Place among the brands named

Why results vary

AI answers change from one ask to the next, even when nothing changed on your site. The same question can name your store on Monday and leave it out on Tuesday. Engines also change how they answer, for everyone at once, when they update a model or show AI Overviews on more searches. Your own figures move when you add, remove or reword questions.

So NameIO never reads a single answer or a single question as a verdict. It reads many questions over several weeks, it compares the same questions with themselves, and it checks every change against how much your figures usually move from week to week.

Ranges and minimums

A percentage comes with its likely range: the values the true figure probably lies between, given how many answers there are and how much they agree. Answers to the same question tend to agree, so 12 answers to one question tell you less than 12 answers to 12 questions. NameIO counts them as fewer effective answers.

With few effective answers a step shows counts, such as "Named in 5 of 12 answers", not a percentage. With more, it shows a percentage with its range as a rough guide; with more again, in full. The numbers are in Every rule and its value. Adding questions to a step narrows its range more than more weeks do.

The smallest change a step can show

The smallest change, in points of Named, that each comparison shows 4 times in 5, for a step with 7, 15 or 27 questions that don't name your store. Calculated at a level of 50%, where answers vary most, and a correlation of 0.5 between answers to the same question (the starting value until your own is measured). A change smaller than the floor of 10 points is never called, so no cell shows less. "Weeks after a change" counts the weeks until the comparison covers the change, plus one week for an Up or Down on the last 4 weeks to be confirmed.

All 3 engines together:

ComparisonAnswers a weekWeeks after a change7 questions15 questions27 questions
Last 4 weeks vs the 4 before15181110 (floor)
Last 4 weeks vs the 4 before351110 (floor)10 (floor)
This quarter vs the one before (13 weeks each)1131010 (floor)10 (floor)
This quarter vs the one before (13 weeks each)31310 (floor)10 (floor)10 (floor)
One check vs the 4 weeks before11291813
One check vs the 4 weeks before31171010 (floor)

One engine:

ComparisonAnswers a weekWeeks after a change7 questions15 questions27 questions
Last 4 weeks vs the 4 before15321914
Last 4 weeks vs the 4 before35181110 (floor)
This quarter vs the one before (13 weeks each)113181110 (floor)
This quarter vs the one before (13 weeks each)3131010 (floor)10 (floor)
One check vs the 4 weeks before11503122
One check vs the 4 weeks before31291813

These figures leave out the swings that move every question at once, such as an engine update, so real figures are larger. When a change can't be told yet, the app gives the figure for your own questions: "a change smaller than about … points can't be told from normal variation".

How a change is called

Each step compares the last 4 weeks with the 4 weeks before, on the questions asked in both, in points, with its likely range. NameIO calls a change only when all of these hold:

What the app then says:

Three more comparisons are read again every week, at a stricter level: since you started tracking these questions, since your latest change in this step, and this quarter vs the one before, once there are two quarters of weeks. A confirmed change leaves a mark on the chart, which stays after the 4-week comparison goes back to steady.

Fixes. A fix gets one reading, at Day 28. NameIO compares the questions the fix touched with your untouched questions over the same weeks and engines, so a change that moved every question doesn't count for the fix. The reading is then kept as it is; a new method version can restate it, with the day. None of these readings says why a figure moved.

How often a change is called by chance

NameIO tests its rules on simulated stores where nothing changed, so every change it calls there is a false one:

Measured on Nov 11, 2026 with the rules in use, on simulated stores where nothing really changed, so every change called is a false one. Each store was simulated 500 times, each time for a year of weekly checks on 3 engines.

Simulated storeQuestionsAnswers a weekYears with a false Up or Down
Named in about half the answers, correlation 0.42730.0%
Named in about half the answers, correlation 0.42711.2%
Named in about half the answers, correlation 0.62730.0%
Named in about half the answers, correlation 0.62710.4%
Matched to a real store's week-to-week flips, correlation 0.42730.0%
Matched to a real store's week-to-week flips, correlation 0.42710.0%
Matched to a real store's week-to-week flips, correlation 0.62730.0%
Matched to a real store's week-to-week flips, correlation 0.62710.2%
A small awareness step734.2%
A small awareness step710.0%

How far the rules agree with a person

Steps and place come from word rules. To check them, a person labels the same questions and answer pairs once, without seeing what the rules said, and the two are compared:

Stage rules (version 1, questions of Nov 8, 2026). Not yet measured: this fixture is waiting to be labelled. The rules are unsure (low confidence) on 8 of its 150 questions.

Wording around brand names (word list version 4). Not yet measured: this fixture is waiting to be labelled. Until it passes, the rank is shown as a rough guide.

Hidden or partial data

Some weeks hold less than a full picture, and NameIO says so instead of filling the gap:

Limits

Every rule and its value

These are the rules of method version 3, with the same values since version 3 (Nov 14, 2026).

GroupRuleValueWhat it means
LevelMIN_EFFECTIVE30A step shows a percentage only from this many effective answers; below that it shows counts.
LevelFIRM_EFFECTIVE80From this many effective answers a percentage is shown in full, not as a rough guide.
LevelMIN_CLUSTERS5A percentage also needs answers to at least this many different questions.
LevelDEFAULT_RHO0.5How alike answers to the same question are taken to be (0 not at all, 1 always the same) until your own answers can measure it.
LevelRHO_PRIOR_WEIGHT20When your own measure is blended with that starting value, the starting value counts as much as this many questions' answers.
LevelRHO_MIN0.05The blended value is kept at this or more.
LevelRHO_MAX0.95The blended value is kept at this or less.
ChangeALPHA0.05The chance of calling a change that isn't there, accepted for one test.
ChangeALPHA_SINGLE0.01The stricter chance used while an engine gives one answer per question a week.
ChangeANCHOR_ALPHA0.01The stricter chance for comparisons read again every week: since you started, since your latest change, and this quarter.
ChangeMIN_CHANGE_CLUSTERS6A change needs at least this many questions asked in both periods.
ChangeEXACT_BELOW15Below this many questions, the test counts every possible pattern instead of using an approximation.
ChangeWINDOW_WEEKS4One period is this many counted weeks.
ChangeQUARTER_WEEKS13A quarter is this many counted weeks.
ChangeMIN_COUNTED_WEEKS3A period needs at least this many counted weeks.
ChangeCONFIRM_WEEKS2An Up or Down is confirmed, or cleared, by this many weekly readings in a row.
ChangeBASKET_OVERLAP0.8At least this share of each period's questions must be asked in both periods.
ChangePANEL_BREAK0.1A line splits where more than this share of a step's questions changed.
ChangeSAME_BASKET_SHARE0.75The same-questions line keeps the questions asked in at least this share of the weeks shown.
ChangePRESENCE_FLOOR_MIN5The smallest floor for Named, in points. A move under the floor is never called a change.
ChangePRESENCE_FLOOR_MAX10The largest floor for Named, in points. The floor is half the room left to move, kept between the two.
ChangeSHARE_FLOOR_POINTS4The smallest floor for share of brand mentions, in points.
ChangeSHARE_FLOOR_SHARE0.15The floor for share of brand mentions is at least this part of the earlier share.
ChangeNEW_ENGINE_LEVEL_WEEKS4A new engine joins the all-engines level after this many counted weeks.
ChangeNEW_ENGINE_TEST_WEEKS8A new engine joins the all-engines change after this many counted weeks.
WatchEWMA_LAMBDA0.2How much weight the newest week gets in the moving average behind Watch.
WatchEWMA_L3How wide the Watch limits are, in multiples of the usual week-to-week spread.
WatchEWMA_BASELINE_WEEKS8Counted weeks collected after each restart before Watch starts.
FixSPREAD_CHECKS4Before a fix: at most this many counted checks.
FixSPREAD_DAYS35Before a fix: checks from this many days before it was added.
FixAFTER_CHECKS4After a fix: at most this many counted checks.
FixMIN_AFTER_CHECKS3A fix is read only with at least this many counted checks after it.
FixFIX_NAMED_FROM0.5A question on one engine counts as named in the check that found the gap when at least this share of its answers named your store.
FixMIN_FIX_CELLS6A fix needs at least this many question and engine pairs that changed.
FixMIN_CONTROL_INTENTS10A fix is compared with your untouched questions only when there are at least this many.
FixPLACEBO_DRAWS1000Random redraws behind a fix's comparison with untouched questions and behind the range of your place.
FixFIX_ROLLUP_Q0.1When fixes are counted together, the share of false calls allowed among those called.
MoversMOVERS_Q0.1In the list of questions that moved most, the share of false calls allowed.
MoversMOVERS_POINTS20A question is listed only when it changed by at least this many points.
MoversMOVERS_MAX_ROWS5The list shows at most this many questions.
MoversMOVERS_CONCENTRATION_ROWS4When at least this many listed questions share one engine and one direction, the list says most of its moves are on that engine, in one direction.
RankMIN_DECISIVE30A place needs at least this many decisive answers.
RankRANK_FIELD_MIN5A brand joins the field of your place when it is ranked in at least this many answers, or in the share below of the step's ranked answers if that is more.
RankRANK_FIELD_SHARE0.1That share of the step's ranked answers.
RankRANK_FIELD_OTHERS9At most this many other brands are in the field.
Data qualityMIN_COVERAGE0.9A check counts as a week only when it collected at least this share of the questions asked on that engine.
Data qualityTRIGGER_EVENT_POINTS15A move of this many points in how often Google shows an AI Overview is marked on the chart.
Data qualityCACHE_RATE0.1When more than this share of answers in a check are exactly the same, the app flags that answers may be repeats, not new asks.
Data qualityFORMAT_RECALL_DROP10A drop of this many points in how many brands are found in one answer format is marked on the chart.
Site checksSMALL_SAMPLE10Site checks and older readings: a change is shown only when both readings count at least this many.
Site checksFLAT_POINTS2Site checks and older readings: a change under this many points is shown as no change.
Site checksFLAT_COUNT2Site checks and older readings: a change under this many answers is shown as no change.
NamesNAME_LABEL_MAX_WORDS4A lead of at most this many words can be read as a section label, not a name.
NamesNAME_NOUN_MAX_WORDS3Only names of at most this many words are tested as everyday words.
NamesNAME_HEAD_MAX_WORDS2The first part two names share, or a known brand's name inside a longer one, is at most this many words.
NamesNAME_MAX_HOSTS3The most linked sites kept for each name.
NamesNAME_ADJ_MIN_HITS3A name counts as a brand's product when answers write it right after that brand at least this many times, and the next two rules hold.
NamesNAME_ADJ_MIN_SHARE0.1Those times are at least this share of the times the name is written.
NamesNAME_ADJ_OTHER_MAX0.2No other brand is written right before the name in more than this share of those times.
NamesNAME_CARD_ADJ_MIN2A name also counts as a brand's product when at least this many product card titles open with that brand and go on with the name.
NamesNAME_LOWER_MIN2A name answers write in lower case mid-sentence at least this many times can be an everyday word, not a name.
NamesNAME_LOWER_RATIO10It is one when it is written in lower case at least this many times as often as with a capital.
NamesNAME_WORDS_MAX_NAMES100The most names read in one answer for the words written around them.
NamesNAME_WORDS_MAX_SPELLINGS200The most brand spellings looked for right before a name in one answer.
NamesNAME_WORDS_MAX_AFTER5The most brands kept for each name as the ones written right before it.
NamesNAME_CARDS_MAX50000The most product cards read when names are sorted, the newest first.
NamesNAME_LIST_EVENT_NAMES3The brand list's change log gets an entry when at least this many names move in or out of the count.
NamesNAME_LIST_EVENT_SHARE0.05It also gets one when the names that moved held at least this share of the brand mentions counted.
NamesNAME_C2_ON11 when a name sold as a product on your store's site or a tracked brand's site counts as that brand; 0 when off.
NamesNAME_C3_ON11 when a name listed as the seller on product cards is a retailer; 0 when off.
NamesNAME_A2_ON11 when product card titles can make a name a brand's product; 0 when off.
NamesNAME_F2_ON11 when names that share a first word answers never write in lower case can count as one brand; 0 when off.

Changelog

Each method version is dated the day it took effect. A version that counts history again says so, with the figures that moved; a change mark or a Day 28 fix reading that comes out differently is shown as restated, with the day.

Version 3, Nov 14, 2026

Version 2, Nov 9, 2026

Version 1, Nov 8, 2026