Your product page
- Weight limit
- 3.5–15 kg
- Newborn insert
- Included
- Last updated
- This season
Measure first. Everything else is opinion.
A week of live readings before anything changes, and the questions worth working on chosen out of that week with you rather than handed to you. Everything below follows from it — what the data says is wrong, what SearchD ships, and how you tell next month's movement from next month's noise.
The calendar
How long does this take?
Four windows, and they do not run on one clock. What SearchD ships on your own site is scoped in weeks; the first re-measurement is scoped in months, because an engine has to re-crawl and re-answer before any of the work can show. These are the windows an engagement is scoped against, not a promised date for a result.
Week 1
1 week
Select
Around fifty candidate questions, the fan-out queries engines generate for themselves included, tracked five consecutive days. That week sets your noise floor and is the evidence you and SearchD read together to choose what to concentrate on.
Week 1–2
1–2 weeks
Publish
The pages that answer your buyers’ questions, and the fixes that let an engine read them. None of it waits on another party, which is why it is the part scoped in weeks rather than months.
Month 1–2
30–60 days
Re-measure
The same questions again, on the same protocol. An engine has to re-crawl and re-answer before any of the work can show, so this — not the week it ships — is the window we scope first movement against.
Month 3–6
3–6 months
Corroborate
Getting the facts right where the model reads, and keeping the panel moving after the first round. Slower, because it needs other people to publish or correct something.
Windows we scope engagements against, not measured averages.
The first week
Who decides which questions you measure us on?
You do, with SearchD, after a week of data rather than before it. We assemble around fifty candidate questions — including the fan-out queries the engines generate for themselves — and track them for five consecutive days. Then we sit down with you and the same report and pick the three to five the work will concentrate on.
01
~50candidates
Assemble the candidates
Questions your buyers actually type, drawn from your Search Console query strings, your reviews and your support tickets — plus the fan-out queries the engines generate for themselves, which never appear in a keyword tool.
02
5days
Track them unchanged
One week, same conditions, nothing altered. The week does two jobs: the spread across it is your noise floor, and the runs themselves are the evidence the next step is decided on.
03
3–5to concentrate on
Read the report together
You and SearchD sit with the same week of data and choose what the work concentrates on. You also agree which number the engagement is judged by. Nothing is handed to you to approve after the fact.
What the week's runs are read for
- Does the engine cite sources for it at all?A question answered from memory, with no sources under it, cannot be moved by publishing anything. It comes out.
- Are the cited sources the same run to run?The same handful of domains repeating means there is a reading list to get into. A different set every run means there is not.
- How crowded is the answer?A few names competing leaves room. A long list, or the same one or two names every time, does not.
How big is the panel, then?
It is an output of that week, not a number SearchD brings to it. How many questions keep being measured and how many the work concentrates on are both set in the session and written into the engagement, because a freight marketplace and a sunscreen brand do not have the same number of questions worth buying. What does not change afterwards is the list: once it is agreed it is frozen, and it is yours — a panel you do not hold is a denominator you cannot audit. A question added mid-engagement has no earlier reading behind it, so it is reported separately rather than folded into the series.
The noise floor
How do you tell a real gain from a number that moved on its own?
By measuring the instrument before measuring the work. Every SearchD engagement will open with those same five days run with nothing changed, and the spread across them becomes your noise floor. SearchD prints every later reading against that band, in both directions.
The stability run — same panel, five consecutive days, nothing changed
Illustrative — not a runn = 30 questionsone engine1 run a day
- 32%Day 1
- 38%Day 2
- 30%Day 3
- 36%Day 4
- 33%Day 5
Your noise floor — 30–38%, 8 points wideShare of the 30 questions tracked whose answer named you, one engine
What a later month's reading means against it
| A later month reads | Against the band | The report says |
|---|---|---|
| 37% | inside the band | noise |
| 31% | inside the band | noise |
| 46% | 8 points clear above it | movement |
A control is what decides whether a published AEO result means anything. When 1,885 pages added schema markup, citations rose 43% — until the same window was measured on matched pages that changed nothing. After that control the effect was +2.4% on Google AI Mode and +2.2% on ChatGPT, neither distinguishable from zero, and −4.6% on Google AI Overviews (Ahrefs, 11 May 2026). A noise floor is the same instinct applied to your own number rather than to a study. The evidence page works through what else SearchD checked before selling it, and what did not survive.
The number
Mention, citation or sentiment — which one are we buying?
SearchD measures all three and you agree one of them, in the same session that picks the focus questions, as the number the engagement is judged by. They are not three views of one dashboard: each one sends the work somewhere different, so choosing between them changes the plan rather than the chart.
| Axis | What a run counts | What the work becomes | A win looks like |
|---|---|---|---|
| Mentionnamed rate | Whether the answer names you at all, as a share of the questions tracked. | Pages that answer the question outright, one resolved entity, and getting into the sources the shortlist is assembled from. | Your name in answers that did not carry it before. |
| Citation | Whether your own page is among the sources the answer was built from. | Publishing your specification as the primary source worth quoting, structured data, and rewriting passages so one lifts whole. | Your domain in the answer's source list, not just your name in its text. |
| Sentiment | How the answer characterises you on the occasions it does name you. | Correcting the record at whichever source the characterisation is coming from, and the review and community work behind it. | The same mention, described accurately. |
Which number is the guarantee written against?
None of these three. The guarantee is written against organic traffic: US Google clicks that do not search your brand name, plus US visits from AI assistants, read from your own Search Console and GA4 and adjusted for season. It stays there whichever axis you choose to optimise, and the guarantee page gives the formula. Named rate is still how SearchD reports AI visibility inside the work, so the term in the report and the term on this page are the same term.
The two audits
Why does the audit run as two separate passes?
Because the pages you own and the sources you do not are checked by different methods, return different findings, and move on different clocks. SearchD runs the on-page pass against your own property, where an approved fix ships the same week; the off-site pass reads the sources the engines cite for your category, where a correction waits on somebody else.
| Dimension | On your property | Off your property |
|---|---|---|
| What it establishes | Whether an engine can reach your pages, read them, and lift a passage out of one. | Which sources the engines cite for your category, and what those sources say about you. |
| How it is checked | Fetched from outside your network as each named crawler, read as raw HTML, passages tested by extraction. | Read off the citation set your own panel recorded — every source behind every answer, run after run. |
| What comes back | A cause per gap, with the page and the line it sits on. | A ranked source list, marked where you are absent and where you are present but wrong. |
| Which of the five causes | 01 – 03 | 04 – 05 |
| Who has to agree | You. | An editor, a moderator, or whoever controls the listing. |
| How fast a fix lands | Days | Months |
Five causes, cheapest to rule out first
The five divide across the two passes, and the order is the argument: a diagnosis that starts at the bottom spends the budget before it reaches the answer. The first two are settled on your own property in under half an hour. Entity resolution starts there and finishes off it. The last two are not yours to fix alone.
| Order | Cause we check | Time to rule out |
|---|---|---|
| 01 | Crawlers blockedA bot rule returns 403, or the page needs JavaScript to show its text. | 5 min |
| 02 | Nothing quotableReadable, but no paragraph survives being pulled out of the page. | 15 min |
| 03 | Entity unresolvedYour name exists in several forms. None of them wins. | 1 hour |
| 04 | Nobody corroboratesOnly your own site states your specification. | half a day |
| 05 | Competitor owns corpusYou are absent from the reading list, not from the market. | days |
On your property
What do you check on the pages we control?
Five things, in the order they can be ruled out. SearchD fetches your pages from outside your network as each named crawler, reads what comes back as raw HTML, and tests whether a passage survives being lifted out of the page it sits on. Then the entity strings and the schema, and last the facts themselves.
- Crawler access, verified from outside your network. A
robots.txtallowing GPTBot proves nothing if an edge rule challenges it, so each agent is fetched rather than read about in config. - Raw-HTML completeness. Whether the page is whole before any JavaScript runs, because the renderers that matter here do not reliably run it.
- Passage lift. Whether a paragraph still names you and still carries a number once it is pulled out of the page around it.
- Entity strings and schema. One spelling of your name everywhere, resolved to one node, so the evidence for you pools instead of splitting.
- Factual accuracy against your own specification. Every number an engine states about your product, checked against the document you signed rather than against your marketing copy.
What the third check actually looks like
Cannot be lifted
“This is why it matters so much here, and it is worth understanding the reasoning before you decide.”
No subject. No number. Pulled out of the page, it names nobody. Sample copy, fictional brand — the shape is the point, not the words.
Can be lifted
“Haruwell Daily Fluid is SPF 50+ PA++++, uses chemical filters only, and was tested at the full 2 mg/cm² application rate.”
Names the product, carries three checkable numbers, states the test condition. Sample copy, fictional brand — the values are invented.
Separately from whether you are named: what is being said. A wrong SPF value, a wrong weight limit, a wrong age range — stated with the same confidence as a right one. For a regulated product that lands on your compliance owner's desk. The same defect in a K-beauty account is worked through on the category page.
Off your property
How do you find the sources an engine actually reads about us?
From your own runs. Every run SearchD makes records not just whether you were named but every page the answer was built from, so the accumulated citation set is your category's reading list — measured, not guessed. The off-site audit is that list, classified and marked against your brand.
The source map, and why it is also the target list
Every page the focus questions were answered from is sorted into five kinds, and each one is marked for whether you appear on it:
- Directories and comparison sites — listings you can be added to
- Listicles and round-ups — "best X" articles an editor controls
- Communities — the threads your buyers actually read
- Trade press — the publications your category is covered in
- Competitor-owned pages — rivals' own sites, cited as sources about the category
That table is the off-site work order. Nothing is pitched, placed or bought outside it, because a domain the engines never cite cannot move a number no matter what it costs.
Two failure modes come out of it, and they need different work
Absent is a corpus problem: the source exists, the engines read it, your brand is not on it. Present but wrong is a correction problem, and it is the more urgent of the two — an engine repeating a wrong weight limit states it with exactly the confidence it states a right one.
Seven surfaces one answer is assembled fromsample
- Your siteyours to write
- Amazon listingyours to correct
- Retailer pagesyours to correct
- Review round-upssomeone else's
- Reddit threadssomeone else's
- YouTube reviewssomeone else's
- Home-market reviewswritten, wrong language
How many you write
1/7is a page you controlThe other six decide the shortlist.
Distributor’s marketplace listing cited
- Weight limit
- 3.5–15 lb
- Newborn insert
- Sold separately
- Last updated
- Two versions ago
What the assistant then states“It holds 3.5 to 15 pounds, and the newborn insert is a separate purchase.”
The work
What does SearchD change, and what has to wait on someone else?
Everything on your own property SearchD writes and ships itself, which is why that half is scoped in weeks — it waits on nobody but your approval. Everything off it needs another party to publish or correct something, which is why that half is scoped in months. Whether either moved your number is the next run's answer, not this stage's.
Live on your site, in weeks
- New pages for the questions you are losing
- Your specification published as quotable fact
- Existing pages rewritten so a passage lifts whole
- One entity string everywhere, resolved in schema
- Every page complete in raw HTML
- AI crawlers unblocked, confirmed from outside your network
Off your site, in months
Correct the record
Brand registry where you own the brand, documented correction requests citing your own specification page where a distributor controls the listing. Re-checked monthly, because a corrected listing can revert.
Get into the reading list
The comparison pages, round-ups and communities your category is actually answered from — the ones on your source map, and only those. Earning a mention there needs an editor or a moderator to agree. A claim only you make rests on your word; a claim independent sources also make does not.
Search and AI
Does SearchD work on Google search or on AI answers?
Both, because they read the same sources: SEO for Google, AEO and GEO for AI answers. Google ranks pages and AI assistants quote them, and an answer is assembled from six kinds of source. SearchD works all six, so one piece of work moves your Google clicks and your AI visits together, and the guarantee counts both.
Three of the six need nothing from you at all.
01
Your own pages
Make the passages liftable
Merge the patch
02
Home-market press
Restate it on US domains
Nothing from you
03
The missing answer
Write it and publish it
Sign off the copy
04
Reddit & forums
Show up where engines read
Nothing from you
05
“Best X” round-ups
Get you into the lists
Nothing from you
06
Directories & ratings
Correct the public record
Confirm the facts
Operational issues we resolve
Which errors keep international brands out of US search and AI answers?
Missing mentions are rarely a mystery. They trace to specific conflicts between what Google and AI engines read and what your brand actually sells in the US.
Retail & Store Availability
Sold at Target, Ulta, or Amazon US, but AI answers claim your brand is “not available in the US” or cite discontinued product lines.
Packaging & US Regulations
AI confuses imported packaging with US FDA labeling compliance, misquotes SPF/ingredients, or says orders ship from overseas instead of US fulfillment.
Reviews & Media Evidence
Thousands of home-market reviews (Hwahae and Naver, for a Korean brand), local virality and awards exist, but US AI assistants cite competitor listicles and claim zero US track record.
Pre-Launch & Category Entry
Preparing to launch an imported sunscreen, cushion foundation, or haircare brand in the US next quarter and needing baseline visibility before retail entry.
Competitor Category Dominance
Competitors dominate every prompt for sunscreens, serums, or sheet masks while your authentic product is bypassed in English answers.
Entity & Distributor Attribution
Former US distributors or same-name entities take credit for your products, or AI answers fail to connect the overseas founder to the US brand.
What you hold
What do we actually receive?
The method's own record, in a form your team can act on without us. Every observation SearchD takes is written down and handed over, and it stays yours whether or not you continue. If you never work with SearchD again, your existing agency can act on the lot.
| What you keep | What is in it | Where it comes from |
|---|---|---|
| The observation record | Question, engine, date, model version, whether you were named, the sentence verbatim, who was named instead, and every source the answer cited. | Every run |
| Your noise floor | The five-day spread from the first week, per engine, and the band every later reading is printed against. | Week one |
| The selection record | Every candidate, what its week of runs showed, and why each one was kept or cut — so the list can be argued with rather than accepted. | Week one |
| A cause per gap, with its evidence | Which of the five causes applies, the page or the source it sits on, and what ruling it out cost. | The on-page pass |
| The source map | Every page the focus questions were answered from, classified, and marked for whether you appear on it. | The off-site pass |
| The factual accuracy report | Every number an engine states about your product, checked against the specification you signed, with the URL it came from. | Both passes |
How those are packaged, scheduled and priced differs by engagement, and this page does not restate it. The audit page lists what a fixed-price four-week diagnostic delivers; the retainer page lists what a month of it adds. One deliverable spelled two ways is the drift this method exists to remove, so it is spelled once, where it is priced.
Every month
How do we know any of it is working?
SearchD re-runs the same questions on the same protocol, several times across the month rather than once, and reports the average per engine against the band set in week one. A single reading is not a result — the same question asked twice does not reliably come back the same way, which is the whole reason the band exists.
- 01Re-run the panelThe same questions on the same protocol, three runs across the month.
- 02Ship on-site2–4 pieces, weighted to the questions you are losing.
- 03Work the corpusCorrect the record where retailers and marketplaces have it wrong.
- 04Report one numberThe agreed metric per engine, averaged across the month’s runs, every delta printed against your noise band.
The shape of a month under retainer. No engagement has run a full cycle yet — when one has, the results page is where it shows up.
Every month, against the same baseline
Why the target is a few questions and not all of them
Because the arithmetic says so. Even at an 80% chance of appearing on any one question, the chance of holding all ten of a ten-question set at once is about 11% — so a target set at every question is a target nobody can honestly sign. SearchD names a number of questions it expects to move, not a number of answers it promises to be in. A question already performing at the week-one baseline does not count toward it either, because clearing a bar you were already over is a base effect rather than a result.
And if the traffic does not move
The measurement is only worth something if it is allowed to return a verdict against us, so it is attached to the invoice. For a qualified brand, if US organic traffic is not up 30% at day 90, adjusted for season, there is no management fee until it is. The full terms cover the formula and what moves the dates.
The guarantee, in four dates
- Day 0You give access. We freeze and hash your 91-day baseline.
- Day 63–90The measurement window, compared with the baseline and adjusted for season.
- Day 97We send the scorecard. You do not have to ask.
- Below 30%?No management fee. We keep working until you reach it.
Qualified brands, first 90 days of the first engagement. Full terms
Method questions
Answers vary every time you ask. How is this a measurement?
They do, so SearchD controls what can be controlled — same conditions, clean sessions, no personalisation, region pinned, model version recorded on every observation — and treats the rest as real variance rather than pretending it away.
The rest is what the first week is for: it sizes that variance before any work starts. A month is then reported as an average across several runs rather than a single reading, per engine, with that band printed beside it.
What if a model update moves the number instead of your work?
It will, in both directions, and nobody can fully separate the two. SearchD records the model version on every observation, re-runs within a week of a major release, and reports against your noise band. Where a jump is plainly a model update, the report says so — an agency claiming every rise as its own is telling you it is not measuring carefully.
Why does off-site work matter more here than in SEO?
A model is asserting a fact, not returning a link. A specification only your own site states is a claim it may decline to repeat, because the only source for it is the party who benefits from it. Independent agreement is what removes that problem.
How much of our team's time does this take?
The session where the focus questions are chosen, access to analytics and Search Console, and one round of review on anything SearchD writes for you — about 90 minutes in total for the audit. We do not run a content calendar you have to staff, and nothing ships without your approval.
See who US buyers find instead of you.
Send us your URL. We run 20 buyer questions across five engines by hand and send you the transcript back within 48 hours.