What AI says about you.
Why your products don’t show up.
What we do about it.
We ask five AI assistants the questions your buyers actually ask, and keep every answer. We name the one reason you were left out. And we go through your product data field by field, and check whether AI shopping results can read your products at all.
AI shopping removes you before it ranks you.
Ask an assistant for something under a budget, in stock, delivered this week, and it throws out everything that does not match before it decides an order at all. An item whose price it cannot read is not ranked lower. It is dropped, and nobody sees it happen.
Leave one of these blank and the item is thrown out. It does not appear lower down. It does not appear.
Leave one of these blank and you are still in the running, further down.
We check both on every item, not on your product data as a whole. A field that fewer than 90% of your items carry counts as missing, because AI shopping does not stop and ask.
| What is missing | What it costs | Whose morning |
|---|---|---|
| One of the 7 fields that get you removed | The sale. The item never appears. | A merchandiser, usually one column in a spreadsheet |
| One of the 8 that decide position | A worse place among the products you are shown beside | Whoever writes your titles and descriptions |
| Nothing. Everything is there and a rival was picked | Nothing in your product data is wrong | Reputation and third-party coverage, over months |
Telling those three apart is the whole job. A tool that only ever reports a fault is a tool that finds one whether or not there is one, and the third row is the finding we are most often able to make.
No vendor publishes the filtering rule and we are not claiming one does. A filter removes what does not match before a ranker orders what is left, which is what makes an unstated field an exclusion rather than a penalty. The field split is ours, and it is the one Antahe checks. See /agent-ready for every field.And when the sale clears, it is bigger. Shopify’s own Q1 2026 merchant data puts AI-referred orders 14% higher in average order value than orders from organic search, growing nearly 13 times year over year. Adobe measured AI-referred traffic converting 42% better with 12% higher engagement. Salesforce credits AI and agents with $262 billion of the $1.29 trillion in global online sales over the 2025 holiday season, one dollar in five.
Shopify, Q1 2026 merchant commerce data, published 11 May 2026. Adobe Analytics, March 2026 reporting period: the same measure has ranged from minus 9% to plus 60% across their previous twelve months, so read it as one period rather than as a constant. Salesforce, 2025 holiday shopping data, 1 November to 31 December 2025, global online sales.Four layers. Each answers one question.
They stack. You cannot explain what you never measured, and you cannot fix a problem nobody has named. The bottom two are where the money is, and they get the room here to match.
Presence
Are you in the answer at all?
We put the questions your buyers actually ask to five AI assistants, on a schedule. Five wordings of each question, each asked once, so one lucky answer never carries a number.
- The questions your buyers actually ask, asked five ways on five assistants
- Every answer kept, word for word
- A range on every number, never one good day
Under the hood
- Prompt harness
- Source landscape
- Wilson intervals
Belief
What does it say about you?
What an assistant states about you, the one word it has attached to your name, and which of your buyers hears which version. Every answer is kept, so you can argue with a claim.
- The one word AI has attached to your name
- Which of your buyers hears which version
- Every claim traced back to the answer it came from
Under the hood
- Association map
- Claim extraction
- Digital twin crawler
- Fan-out reconstructor
Commerce
Can AI shopping buy from you?
We read your product data and check every item against the fields AI shopping filters on. A field that fewer than 90% of your items carry counts as missing, because AI shopping does not stop and ask. Where your product data is clean and you still were not picked, we say so, which is the finding a tool that only reports faults can never make.
- Your product data checked field by field, item by item
- The one field that removed an item, named
- Fixes shipped to your site and your product data, with your approval
Under the hood
- Catalogue upload, CSV or TSV
- Product page reader, no connection needed
- Per item field completeness
- Price and stock mismatch
- UCP manifest probe
Facts
Can every AI get your facts right?
One master copy of the facts about you: what you sell, what it costs, what you claim. When an assistant repeats something you stopped publishing months ago, we catch it against that copy and show you the answer it came from. Correct a fact once and every later check reads the corrected one.
- One master copy of your facts, in one place your team edits
- An assistant repeating something out of date gets caught against it
- A claim with no fact on file is skipped, never guessed at
- Serving those facts to an AI on request specified, not built
Under the hood
- Canonical fact store
- Hallucination watch
- Stale fact diagnosis
What you actually log into.
Two screens, from the two halves of the product. The product data one is first, because it is the one that decides whether AI shopping can buy from you.
Nine screens open off it, built around one table. Every number links to the answers behind it, every finding becomes a job of work carrying the date we re-checked it, and a number we could not measure says so rather than showing a zero.
What AI believes your products can do
When a buyer names a constraint, whether your catalogue could answer it and whether you were picked
Product data
Every item against every field AI shopping filters on, and which field is getting items removed
The evidence a crawler could read
Your page exactly as an AI crawler received it, step by step, and which of the five gates between it and your words is closed
Plan
Every finding as work, with who it needs, roughly how long, and the arithmetic behind that
Studio
The rebuilt page beside your original, with a line per edit naming the rule that required it
Proof
Three readings taken independently, and what they do not establish
Where you stand
What each assistant says, who is recommended instead, and what every audience hears
The belief, in their own words
Every question we asked and every reply word for word, and every assertion an assistant made about you
Questions
The set we propose, before we spend a rupee collecting it
How a missing answer gets diagnosed.
A merchant, five steps, no gaps. This is the row marked at the top of the page, followed from the question to the re-read. The brand is invented so every step can be shown, including the ones a real client would not want published.
- 01
Someone asks one question.
Not a keyword. A sentence, with a budget and a delivery window buried in it.
- 02
The model turns it into about ten searches.
One after another, each built on the last, and no analytics product shows you any of them, Google's own included. We work out what those searches were and check whether your pages could have answered them. For a product the last one is not a question at all: it is a filter on price, stock and delivery, run against your product data. One decides whether you get named. The other decides whether you can be bought. We check both, because a buyer meets them as one thing.
- 03
We read your product data and your product page.
The page is fetched with no JavaScript, because that is what an AI bot runs. Zero execution was detected across more than 500 million GPTBot fetches. Then the two are compared, field by field, so a price stated in one and not the other is a finding.
Lantern, 2026, more than 500 million GPTBot fetches
- 04
We put the two together and name one reason.
Asking the question tells you that you lost. Put that beside what the AI bot and your product data actually held, and you get one plain reason from a fixed list of twenty: nine about your products, eleven about what AI says.
- 05
We ship what is ours, then re-read.
The same questions under the same conditions. If the change sits inside the interval we say so rather than claiming a win.
1,240 items read · delivery stated on 828 · absent on 412 · 67% coverageGET /p/meridian-18v · user-agent: GPTBot · 200 OK · delivery in the page, not in the feedA rival whose feed carries a delivery window on every item, so all of theirs survived the filter and 412 of yours did not.
Populate the delivery field for the 412, publish it as offer markup on the product pages too, then re-read the same prompts under the same conditions.
Left out of 9 of 12 shopping questions. The delivery field was blank on a third of the products, so those items were never even considered.
One supplemental feed field. Forty minutes of a merchandiser's morning.
Shown in 7 of 12, re-read over 90 days at 25 answers per question per read: five wordings across five assistants, each asked once. A range is reported on every figure.
One prompt went the other way. It is in the report, next to the gains.
Twenty plain reasons. Not a score.
Eleven explain why AI answers leave you out. Nine explain why your products get skipped in AI shopping results, and six of those nine are fixed by correcting product data you already publish, which makes them the fastest money in the set.
Every time you are missing, you get exactly one of these, with the evidence attached. Each one is worked out from things we already hold, so you can check the reasoning yourself.
Nine reasons. Why your products get skipped.
The fastest way back to a sale. Of these nine, six come from comparing two things you already publish, so they cost nothing to find and very little to fix.
Your products got skipped in AI shopping
The fix is usually one field in your feed, which makes these the cheapest findings in the set and the fastest way back to a sale.
The product is in no list AI shopping can read
NOT_IN_FEEDYou sell it and it is in no list an assistant can read, so it cannot be picked at all.
A field an AI shopper filters on is blank
FEED_INCOMPLETEIt is listed, with a field the assistant filters on left blank. Blank means removed, not ranked lower.
Your feed and your page state different prices
PRICE_MISMATCHWe are not judging the price, only reporting that your own two sources disagree.
Your page shows the shopper one price and the machine another
PRICE_SCHEMA_MISMATCHAn assistant quotes the machine-readable one and your buyer reads the page, so the sale goes at the step where they disagree.
Your feed and your page disagree about what is in stock
STOCK_STALEYou are either sending buyers to a dead end or hiding something you could sell today.
The link in your feed does not reach a working page
LINK_BROKENA product an assistant cannot open is a product it cannot buy.
Your listing is written in adjectives
GENERIC_ATTRIBUTESA shopper who names a finish, a size or a material has nothing of yours to match against.
You never use the words your buyers use
SEMANTIC_MISMATCHIt clears every filter and is described in words nobody searches with.
Nothing of yours was wrong. Something else got picked
OUTRANKEDEvery field is there. Nothing in your product data is blocking this one.
Eleven reasons. Why the answer leaves you out.
They never reached you
The fix is plumbing, and it is usually the fastest one we ship.
AI bots are blocked at your door
ACCESS_BLOCKEDNo AI bot got in. We watched for them and saw none, and our own fetch was refused or handed a challenge page.
Your page is unreadable to AI bots
RENDER_INVISIBLEAn AI bot reads far fewer words than a person sees. The page renders for people, not for machines.
AI bots get a different page than people do
CONTENT_MISMATCHSame address, same amount of text, different words. What the bot was served is not the page a person reads.
You are missing from an index several assistants read
NOT_INDEXED_BINGSeveral assistants look you up in an index you are not in.
They reached you and you lost
The fix is the words on the page, and we score the rewrite before you publish it.
No page of yours answers what the buyer asked for
CONSTRAINT_FAILThe answer names something specific. No page of yours satisfies it. A rival's does.
You never wrote the page for that question
ONSITE_ABSENTThe question has an obvious page behind it and you have not written that page.
You are dropped at the final reliability check
WAVE3_ELIMINATEDOne question becomes several searches, one after another. You survive the round that compares options and get dropped when it checks who is reliable.
Nobody else is writing about you
The fix is not on your own site, so you get a ranked list rather than work we ship.
Not one of the sources it read mentions you
OFFSITE_ABSENTIt read a set of pages to build the answer. You are on none of them.
It has you confused with another company
ENTITY_CONFUSEDThe mentions it found belong to a different company with a similar name.
With search switched off, it does not know who you are
PARAMETRIC_GAPAsk the model with no web access and your brand is not in what it learned.
It repeats something you stopped publishing
STALE_FACTA claim it makes about you contradicts what is on your site today.
The first group is more common than anyone expects. Roughly 41% of B2B sites still block at least one major AI bot.
Observed across first audits, and mostly settings left over from 2023. All twenty reasons are worked out from evidence we hold rather than asked of a model, because being able to check them is the point.Four of these run right now, free.
You fix what the audit found, then someone ships a page and breaks three of them again. These four run with no account, and the first one runs here, in this page.
Will AI read this page?
Paste a draft or a link. We read it the way an AI does and tell you what survives.
Can AI read your site?
Which AI bots your server turns away, which of your pages come back blank to one, and whether the Bing index has you.
Run itIs your store ready for AI checkout?
All 15 fields an AI shopper looks for, checked against what you publish.
Run itWhich AI bots visit your site?
One line on your site, and you see which bots came, what they asked for and what they got.
Run itDoes the answer come first
Your definition arrives in the fourth paragraph. An AI reads from the top and stops early.
Can each paragraph stand on its own
Three paragraphs open with “it” and lose their subject the moment one gets quoted alone.
Are there any actual facts in it
Two numbers in 900 words, both rounded. Quotations and figures were the strongest changes tested.
Can an AI bot read the page, if you paste a link
Whether it is let in, what it is served, and what it reads back.
Changing the structure alone improved how often a page got quoted by 17.3%. GEO-SFE, arXiv:2603.29979
Scored against the questions your buyers really ask
Needs your question set. Today it scores against questions guessed from the draft itself.
Checked against your own facts
We cannot check a price nobody has given us.
Compared with the pages that beat you
Needs the sources your category actually gets quoted from.
Whether you cover what buyers ask for
Needs to know which segments and which places you sell to.
Nothing is held back. These four need data we do not have yet, and the audit is where it comes from. The checker says the same thing when you run it, in the same words.
Trackers ask what it said. We also ask what it read.
Asking the question tells you a competitor got recommended. Watching your own site tells you the model fetched their pricing page and never fetched yours. One is a symptom. Two is a diagnosis.
- Against tools that only count mentions: they report the score. We report the reason, with the refused request attached.
- Against tools that only rewrite content: content is one reason out of twenty. We work the other nineteen.
- Against the SEO suites: they give AI a tab.
The limits, before you ask.
Every vendor in this category has these. We are telling you before you buy, which is the only part we control.
The engine's ranker
We can establish that every field an agent filters on is present and that something else was chosen anyway. Why it was chosen is inside a system nobody outside the platform reads, and we will not dress a guess up as a reason.
A catalogue you have not given us
We read the feed you upload, the metrics Merchant Center reports and the product pages you publish. An item that exists in none of those three is an item we cannot check, and we will report it as unchecked rather than as clean.
Real user queries
Nobody outside the platforms has them, ourselves included. Anyone selling you real prompt data is selling you invented questions with a confident label.
Retrieval without citation
If a model reads your page and does not cite it, that read is invisible to us and to everyone else.
Revenue attribution
Zero-click influence cannot be tied to a deal. Cross-session influence usually gets miscredited to branded search.
A guarantee that you get cited
The systems are probabilistic and nobody can honestly promise it. We publish our own prediction accuracy instead.
And this channel will not send you much traffic. Under 1% of visits for most sites, and it still decides who gets picked.
Under 1% of visits is what we see in the measurements available to us. We hold conditions constant for trends and vary them deliberately for diagnosis, and we publish what we cannot measure alongside what we can.Run it on your own brand.
Every brand ends up with one word attached to it: cheapest, safest, outgrown, unproven. Find yours.
Free. No account.