Get AI to recognise your business: the same facts everywhere

3.1 Why this comes before writing pages · The two-hour checklist

What you'll do in this section: confirm the baseline is saved, lay out the route for the identity gate's five fixed moves, and get through the six-part checklist in one sitting, in two hours — then decide whether you can take the 30-minute pass. When you're done, you'll have a first-draft entity-error list, a 14-row fact table, an inventory of the five business profiles, and an exclusions list.

Hard gate: until the baseline is saved, do not change the door, any profile or the facts page. How the baseline is run and saved, including that one set of 36 brand-six runs, is in → General Edition 4.3 Coarse screen and piles, full rounds, and the brand-six baseline.

What gets misstated is usually an old or wrong value AI picked up from a third-party page: price, address, opening hours, registration number, a practitioner's name and credentials, or whether the business has closed. This gate has two features found nowhere else:

• Fast payoff. Of every move in the set, it's the only lever where "once you fix it, AI's wording is often the first thing to change" (observed, 70% confidence; not a promise); writing pages, off-site work and video all pay back on a monthly timescale.
• It bears on the conversion rate. After a buyer gets a shortlist from AI, they circle back to check the brand — search the name, look at the website, check the profiles — before making contact.

The only firm number that tells you whether this gate is done: the number of factual errors, X → Y. Seat counts move with sampling, and "on the list and cited" waits on a third party saying yes; this is the only cell that gives a firm number every month, and you can reproduce it yourself by asking the questions again (→ General Edition 7.1 What a retest produces, and the rulers).

How far the evidence goes for each claim in this chapter is set out in full in → General Edition A.1 Evidence for mechanics, the door and identity. The conclusion in one line: AI's live fetch only reads visible HTML, so every fact in this chapter has to be written as visible HTML, one field per line. When you cite that evidence, copy the qualifier along with it — never copy just the number.

The route: five fixed moves

The order can't be reversed: sameAs is for the knowledge graph, not for live fetch. Write the schema before the visible HTML and you end up with a site that's "fields all present but AI can't read it" (the schema template is sealed once, see → General Edition 2.7 Gates 3 and 4: indexing paths, and JSON-LD sealed once).

The two-hour checklist

The 0–25 minute block is read by the person running probes; the rest is done by the person who edits the site. This part does not run a separate probe: it reads that one set of 36 raw answers from freeze week. A separate quick run is not the same ruler as the 36 runs; if X comes from a quick run, X and Y can't be compared. The 75–95 and 95–110 blocks were already done in the door chapter (2.2, 2.5); here you only check the receipts. Canonical domain means the www and https form matches the official site.

What each line of the six-line door read checks is in → General Edition 2.2 Gate 1 read-only audit: what to check in nosnippet, robots and the WAF. One of them matters for this chapter: if Applebot isn't allowed through, don't claim the Apple Business Connect profile (3.4).

On top of this (not part of the 120 minutes; counted per person): one person page for every named practitioner whose name carries trust, 40 minutes to align + half a day to build the page (3.3). An organisation with ≥3 registered practitioners, or any one holding specialist registration (on sides other than the strictly regulated side: or a third-party credential tier) → this must be done in full.

When a 30-minute pass is enough

How to take the pass: fill in only the 14 hard fields of the fact table, check only that the URL field matches across the five profiles, and move every other step into the monthly review column. Log the decision to take the pass in the first line of that week's work sheet — it's a call you made, not something you missed.

3.2 The fact table and the organisation facts page

What you'll do in this section: first build a 14-row fact table as the single source of facts for the whole project; then, following the blueprint, build an organisation facts page (/facts) that machines can copy whole rows from; on the day you change a price, sync the facts page, the thick pages and the five profiles the same day.

The fact table has four fixed columns: Source / Verified date / Publishable? / Owner.

The /facts page: built for machines to copy whole rows, not for error-proofing

The official website's About page doesn't solve this — an About page is made of adjectives; AI can't copy it, and won't dare to. /facts has to deliver: copy it and it's right, copy it and it's complete, copy one row out and it still reads clearly.

It needs its own URL, has to be in the sitemap, and be linked from the footer on every page. The form has to be clean HTML — a <table> or a <dl> — never an image, never a PDF, never JS-rendered: the mainstream AI crawlers run zero JavaScript (only two exceptions; the mechanism is in → General Edition 2.3 Gate 2: four identities, fetched live (read-only)), and if it can't render, the page might as well not exist. The schema follows the one-time template (organisation type + a person type for each practitioner + sameAs); it's not written or checked page by page — that was already sealed in → General Edition 2.7 Gates 3 and 4: indexing paths, and JSON-LD sealed once and isn't repeated here.

How to handle reviews differs by side; it's written once, in full, in → General Edition 5.3 Quick reference by side (2): testimonials and reviews, comparisons, lists, titles, outbound links, FAQ and captions. Here it's only marked by position: on regulated sides, skip the block entirely.

This page type relies on numbers, not a conclusion paragraph: the measured organisation facts pages had no conclusion paragraph, no date and no byline before the first H2, and still got cited (evidence in → General Edition A.3 Page-type measurements (1): two legs, seven rules, the nine tier-A types). So every hard fact card needs a line below giving the definition, denominator and period; miss one and it fails the six checks for numbers (→ General Edition 5.10 Shared parts for all page types: six checks for numbers, no public price, billing units, dates and schema).

Example (education and training): An organisation's homepage H1 was a slogan; AI didn't copy that line — it copied the results card right after the first screen, citing the numbers directly. The page has two flaws: one card says "100%" with no denominator, so it fails the number check; and fees aren't in the homepage HTML. The format is worth copying (a number card right after the first screen, readable as plain HTML); the execution isn't (no denominator, price not in the HTML).

Write H1 as <Legal name> · Organisation facts file, bare facts, zero adjectives; title can use <Service/subject> <organisation type> Singapore | <one checkable fact>, or <Service> in Singapore — <Organisation name>. Neither one carries a slogan.

Example (education and training): A title that has been cited: "JC H2 Math Tuition Singapore | 90% A/B" — the format is worth copying (service + location + one number), the wording isn't (90% has no definition or denominator).

14-row field table (hard fields)

Each row = field name + value + last verified date, with no narrative sentences mixed into the value — a row has to be copyable on its own.

Example (dental and aesthetics): Write row 12 as a firm sentence like "this item is not payable under a given medical savings scheme", not "partly subsidised" or "depends on the situation".

The price block: fixed on the strict side

The price has to be a fixed final price. The detailed wording for each side, and whether a range may appear, is set out once in → General Edition 5.2 Quick reference by side (1): identify the advertiser first; prices, promotions and freebies, result numbers; it isn't repeated here.

Example (dental and aesthetics): A three-tier price list split into mild / moderate / complex, each tier spelling out the defining condition, the tax-inclusive fixed price, what's included and the exclusions — that's what strict-side wording looks like; the lightly regulated side doesn't have to copy it.

Price sentences come in two kinds: an own sentence gives the full price, the billing unit, the number of sessions and what's included; a market sentence gives a range and the variables that affect price. Whether a market sentence may be written at all, and whether it may sit on the same page as your own price, is set by side in → General Edition 5.2 Quick reference by side (1): identify the advertiser first; prices, promotions and freebies, result numbers.

Example (education and training): Own sentence: "<Centre> charges S$480 per month for 4 group <subject> lessons of 1.5 hours each, including GST and notes." Market sentence: "<subject> tuition in Singapore generally costs S$X–S$Y per month, depending on group size, level, and tutor experience."

Some industries' advertising rules say this directly: when the ad space can't fit it, you must give a prominent link to the specific page on the official website, and the source it points to must not be spoken-only, with no record. /facts and the price guide page therefore have a statute behind them — they aren't a spec we invented (Statute text; example industry is training, rule text in the industry edition's appendix — the training edition isn't out yet, and the Dental & Aesthetics Edition appendix doesn't cover this one).

How /facts and thick pages divide the work

/facts is a source that the density gate draws numbers from, not something the gate checks directly: it isn't bound by the thick-page word count, doesn't count against the intent-cluster quota, and doesn't count toward the page total. Every "checkable fact that can't be found elsewhere" on a thick page must point back to a specific row on /facts and its fact-table row number; one that can't be pointed back doesn't count. It also feeds the source gate: the register links and government pages that rows 2, 3, 4 and 12 already carry (plus, on the side that may link out to manufacturers, row 11's manufacturer product page) are the ready-made source for those 3-plus outbound links on the thick page. /facts is still an advertisement: on regulated sides it still has to pass the pre-publication compliance checklist — banned words, price wording, "free" wording, efficacy and peer comparison, regulatory endorsement — not one item is exempt; the checklist is set out in → General Edition 5.2 Quick reference by side (1): identify the advertiser first; prices, promotions and freebies, result numbers and → General Edition 5.3 Quick reference by side (2): testimonials and reviews, comparisons, lists, titles, outbound links, FAQ and captions.

On the day you change a price, sync /facts, the price table on thick pages and the price field in the five profiles the same day, and send an update letter to any third-party page that already lists your price. If the three don't match, you've smudged a fact you signed off on yourself.

How to check: /facts is live; fetch the page source with an ordinary browser UA and search it: the price string and the registration-number string are both there; price and registration number match the fact table word for word, 100%.

3.3 Person pages and six-trace alignment

What you'll do in this section: build a separate person page for every named professional; make the name, registration number, organisation name and address match word for word across six public traces; and when someone leaves, handle sameAs in one of three ways.

This applies to any role that carries the buyer's trust under a person's name — not just the organisation's founders; front-line service providers need one too.

Example (dental and aesthetics): Every named doctor at the clinic counts, not only the founder.

On the regulated sides, almost every form of persuasive content is banned at the organisation level — but not at the person level. The regulator explicitly allows a practitioner to state credentials, scope of practice, practice arrangements and contact details; this is one of the few sources of persuasive content explicitly allowed on the regulated sides, so on the regulated sides a person page isn't a bonus, it's the main line (Statute text; rule text in the industry edition's appendix (example industry: Dental & Aesthetics Edition, Appendix A)).

🔴 Allowed ≠ unregulated: personal information still has to pass the nine compliance criteria (truthful, accurate, verifiable, not exaggerated, not misleading, not inflammatory, not persuasive in tone, not comparative, not disparaging); and there's a separate ban on laudatory terms from the example industry's regulator (e.g. the Annex A laudatory terms table) — never describe this person-level allowance, in any external document or spoken statement, as "a back door around the organisation-level advertising restrictions." Title formatting must follow the format the regulator sets out, and a title not legally held may not be used; anyone not on the specialist register may not imply specialist status (the notice that sets out title formatting could not be machine-read; confidence in this rule rests on a manual check, and it must be checked against the original, clause by clause, before launch; Original text not obtained).

Pure directory listings are the lowest-citation-density page type in the measured data; the largest single measured effect is the entity-strength multiplier (a single third-party study, a US sample; evidence and qualifiers in → General Edition A.1 Evidence for mechanics, the door and identity) — money spent on directory listings buys nothing, money spent on the entity does.

🔴 The output is a page, not an internal table: pulling this together into an internal credentials table doesn't count as done — AI can't read an internal table, only visible HTML; not one of the five major crawling systems reads JSON-LD or hidden structured markup (evidence in → General Edition A.1 Evidence for mechanics, the door and identity).

Fill in the external-anchor field when it exists; don't invent one when it doesn't — especially the academic identifier and the Wikidata QID; whether to build the latter depends on the threshold in 3.4. Third-party credential tiers give only the tier name and the awarding body — see 3.4. Write scope of practice and service prices the same way as the /facts price block; on regulated sides they likewise go through the compliance checklist once (see 3.2).

Each person gets their own URL (/team/<name>), not a paragraph inside the About page. The form matches /facts: a visible <dl> or <table>, one field per row, every external source a clickable <a href> — never JS-rendered, an image or a PDF. After that, every thick page, every video and every piece of off-site material carries this person's byline, and the byline links back to this page — that's the only move that ties page authority to the person's entity.

Fix title/H1 as <Name>, <registered title>; on regulated sides the title uses only the registered wording, never a self-styled one. On regulated-side professional content pages, write a reviewer line as the first line under the H1: the reviewer's name + degree + month and year, followed by the registration number, linking back to the person page; on other page types treat it as a bonus (the measured data on the byline line is in → General Edition A.3 Page-type measurements (1): two legs, seven rules, the nine tier-A types).

Example (dental and aesthetics): Under the H1 on every medical page, write Medically reviewed by Dr X, <degree> · <month year>, followed by the registration number.

The real mechanism behind "8x entity strength" (a single third-party study, a US sample; qualifiers in Appendix A.1) is not what's written in the schema — it's the same set of facts appearing word for word across several independent third-party pages that can be crawled.

Three strings are three entities, and 8x becomes 1x. So the real lever is the six-trace alignment table below, not the page spec table itself.

The six traces change with the anchor layer

The line that holds across every industry: an inconsistent spelling is the number-one reason an entity gets split into two.

Example (e-commerce): The same brand goes by three different names on three platforms; the machine sees three entities, and none of the three sets of social proof adds to another.

Some industries have explicit rules for how the legal name itself must be written, and a non-compliant spelling also splits the entity in two: the spelling on business profiles, list entries and platform pages has to be changed at the same time, or you break the rule and split the entity at once (Statute text; example industry is training, rule text in the industry edition's appendix — the training edition isn't out yet, and the Dental & Aesthetics Edition appendix doesn't cover this one).

Example (education and training): The registered name may not have "Singapore" added after it; to show the location, the only option is to write "located in Singapore" separately.

When an unrelated namesake entity is polluting your results, add one exclusive definition sentence to the first paragraph of /facts: product name + category noun + legal entity name + location, all in one sentence; then make that sentence the first sentence of the description in all six traces (the lift from this is not measured, but the mechanism is clear).

Example (software): A search for a product's SOC 2 status returns an answer that mixes in an unrelated company with the same name.

How long: 40 minutes to align each person, half a day to build the page. Why it moves rankings: machine entity disambiguation runs on strings matching consistently across multiple sources, not on how many times you repeat it on your own page. How to check: the person page opens on the public internet; a string search of the page source finds the name, the registration number and (on sides other than the strict side) the tier name; all six rows × four columns are filled in and match word for word, and every cell that didn't match has a correction letter already sent, with the date it went out; the byline on thick pages and videos clicks back to this page.

When someone leaves: keep the page, handle sameAs three ways

The page isn't deleted because deletion would break inbound links already built. A sameAs pointing to the official register, the association or an academic identifier stays because it proves who this person is; a sameAs pointing to the new employer's page is deleted because leaving it in place feeds authority to a competitor, and might even let entity resolution assign this page to the new organisation. This treatment isn't measured, but the mechanism holds.

There's no such thing as "301 a byline": a 301 acts on an entire URL; a byline field can't be redirected. Only a profile page with genuinely no remaining value gets 301'd whole to the team directory; every other page just has its current owner and author link changed (not measured).

The six places to finish in one pass, the same day: the status line + last-checked date on the person's own page, the organisation's current-staff list, any field across the five business profiles that carries their name, the industry association or platform profile, list entries (send an update letter), and the opening link on their bylined content (point it to the "formerly" page, without deleting the content itself).

When someone new takes over, any sentence on any page that cites the outgoing person's credentials must be changed the same day, or the old and new versions sit side by side — which is itself a self-inflicted factual error.

Where the person is themselves the anchor, and their leaving would take the whole service line with them, the landing page must not be built under that person's name: the subject has to be the service itself, with the person only a replaceable "current owner" field — otherwise, once they leave, the whole page either gets deleted or turns into a wrong entity (not measured, but the mechanism holds).

Example (education and training): For a one-on-one tutoring course page, the subject is "subject × level", and the tutor is only a "current teacher" field.

3.4 Third-party credential tiers, Wikidata and the five business profiles

What you'll do in this section: go get a third-party credential tier wherever you can — it's the substitute for when you can't ask for reviews; apply for manufacturer-locator listing wherever you're eligible; build a Wikidata entry only once the conditions are met; complete all five business profiles in one pass, with the URL field pointing to the landing page, then spend just 20 minutes a month on the three-state check.

Third-party credential tiers: the substitute when you cannot ask for reviews

On the light and unregulated sides, fill in a verifiable third-party tier wherever one exists, and if it can't be looked up, write the date; on the strict side, never write a manufacturer-granted tier (Conservative line (not statute text); basis in the industry edition's appendix; loosening this needs the client's compliance officer's written reasons) — write only technical-standard certification and registration status.

Three hard rules for how to write it:

1. Write only the tier name and the awarding body. "highest", "only", "leading" or "only X on the whole island" hits the banned-word list directly and is banned regardless of whether it's true.
2. On the strict side, case counts and procedure volumes are deleted by default — there's no confirmation process, and no "usable once confirmed" tier; on the other sides, write only your own absolute numbers, and pass all three gates at once: no adjectives, a ledger record signed by the owner, and never placed next to efficacy claims.
3. Tick each item off against the compliance checklist, with the owner's sign-off.

A written sign-off resolves only the three stop-and-escalate situations; it never turns something banned as Statute text into something publishable: the words in rule 1 can't go live even with a sign-off (the stop-and-escalate rules are in → General Edition 0.4 The three labels for compliance sentences, and the stop-and-escalate rule).

Law firms and similar can cite a third-party ranking as evidence of expertise, but which side it falls under is decided by the industry edition; it must pass that industry's publicity-rules review first, and cannot go live before legal signs off.

How to get a tier: apply to the manufacturer for listing in the official dealer / service-provider locator, using your purchase records as evidence; it's usually free, and one application pays off in two places — an inbound third-party link, and third-party endorsement of the real brand and model of the device you use.

Why it doesn't work the other way round: a strict-side page linking out to a commercial company runs straight into the ethical code's own wording (Statute text; rule text in the industry edition's appendix (example industry: Dental & Aesthetics Edition, Appendix A)); this same direction rule also governs where effort goes in the off-site chapter and the substitute tactics, see → General Edition 6.4 Three source groups, and the Group B actions and → General Edition 8.7 Cheap tactics on the unregulated side, and substitutes on the regulated sides — not repeated here.

Wikidata: required when the conditions are met, not open to all

Wikipedia is the single largest node in ChatGPT's citation set, and Wikidata is its structured anchor (numbers in → General Edition A.1 Evidence for mechanics, the door and identity). Its job isn't to get cited — it's to give the entity a unique ID: however long the sameAs chain gets, if there's no node in the graph to catch it, disambiguation still falls back on string matching.

The common misconception is "anyone can build one, and once it's built, it's there" — not true. Admission requires one of the three, and that one has to be describable using serious, publicly available reference material. A practitioner with only a website and a professional social profile who builds one will get it nominated for deletion by the community — you lose the time and also leave a deletion record in the edit history. With none of the three in hand, wait until off-site bylined contribution lands the first piece of third-party coverage, then build it — that's the interface between bylined contribution and this chapter.

Don't edit Wikipedia's body text to add the organisation's name or an outbound link — a self-citation gets reverted and leaves a record; the risk outweighs the benefit. The only thing worth doing is adding a verifiable source to an existing entry, without adding the organisation's name.

Building the entry takes 30–40 minutes; how to check: the QID opens on the public internet, the entry contains ≥3 external references, and a recheck after 30 days shows it hasn't been nominated for deletion.

The five business profiles: an entry ticket, not a lever

The verdict is fixed: this is an entry ticket, not a lever. Complete it once, then just run the three-state check once a month afterwards — never open an "optimisation" task.

Why the order isn't ranked by platform fame: ChatGPT's local answers run mainly through the first two (70% confidence); on the Google leg, 95.5% of cited domains are exactly the URL registered on the profile (evidence in → General Edition A.1 Evidence for mechanics, the door and identity) — and that's the only field in this whole cell with real leverage. The URL on all five profiles should always point to the landing page, never the homepage, each one matched to the page for that buyer type.

Externally, say only the fixed line "a basic step, done in passing" (→ General Edition B.2 Approved wording and scripts) — never "do this and ChatGPT will recommend us": the mechanism confidence for Foursquare, Apple and Yelp is only 70%, and the researchers themselves say there's no data isolating which profile field contributed to the citation.

The compliance scope uses the same word list as the website pages: every free-text box, business Q&A, business post and photo caption on the five profiles is an advertisement. The two places most often missed are photo captions and business posts. Business post topics are limited to: a change in opening hours, a new service added, a change of address, an update to a practitioner's registration information — never a promotional post, never a case post, never reposting a review. Fill in the price field the way /facts does it, by side (3.2); on the strict side, only the fixed final price.

How to ask for reviews and how to reply differs between the two sides; it's written once, in full, in → General Edition 5.3 Quick reference by side (2): testimonials and reviews, comparisons, lists, titles, outbound links, FAQ and captions. Here, just one line: whether you may actively ask for reviews follows the rule for your side, and the answer is in that section, not repeated here.

Monthly three-state check (about 20 minutes): ① present or not ② is the URL field correct ③ has a fact been changed by a third party or the platform. Tick all three off in the monthly report.

How long: completing it once takes about 2 hours; 20 minutes a month after that. How to check: all five are claimed; name / address / phone match /facts word for word; every URL field points to the landing page's canonical domain; each one's launch date is logged in the outreach ledger and added to the monthly review column.

3.5 The brand six questions, and what to do when you find an error

What you'll do in this section: starting from the 36 brand-six runs in freeze week, count factual errors every month with the same ruler; within 1 working day of finding an error, log it and notify; within 3 working days, fix every source you control and send a correction letter; when any of the five trigger conditions is hit, jump the queue and fix it the same day, without waiting for the monthly report.

What the six questions ask, and how the ruler is made up

The brand six is defined once for the whole book, and every mention of brand questions anywhere else in the book points back to this section — change the composition and you've changed the ruler, and the factual-error counts you get out aren't comparable any more.

Ruler composition (the one definition for the whole book): 6 questions × 3 rounds × 2 engines = 36 runs / month, frozen for the whole quarter. The baseline uses only the full-round set of 36 runs from freeze week (4.3), never a quick run — two sets with different compositions aren't the same ruler, and the numbers that come out aren't comparable.

The two Chinese questions run on a separate ruler, saved to their own file, and are never merged into the brand-six number: merge them and you get 48, but only someone with a Chinese page can ever produce 48 — the same "factual errors" count would instantly stop being comparable between businesses with and without a Chinese page. Both numbers go into acceptance, but each keeps its own numerator and denominator. Print the two numbers separately: "Brand six: factual errors X → Y (denominator 36)" and "Two Chinese questions: factual errors X → Y (denominator 12) · wrong-anchor N". Wrong-anchor logs whether the answer quoted a price from the wrong region, or treated an out-of-town business as a local one.

Log four dimensions, read by a person, no skipping

How long: running the full 36 takes about 1 hour; reading takes a person 3 hours. How to check: every raw answer is saved; every error in entity_errors.md has a screenshot of the original text + a source URL + the matching fact-table row number; not one is marked "unread". Why it moves rankings: it doesn't, by itself — it's this chapter's only ruler, and the only signal that can prove both legs move once the source is fixed. AI's factual errors are copied from the source pages it retrieves; fix the source, and both legs move.

After you find an error: the correction path

Whether and when a third party acts on it is their call, not a promise you make; what you can promise is two follow-ups, checking every month whether the error is still there, and re-checking AI's wording question by question. How to write the correction letter and the template wording is in → General Edition B.2 Approved wording and scripts; here, just one line: point out only the error, attach a source link to /facts, never mention rankings, never slip in promotion.

Precondition for sending a letter (a hard gate on regulated sides): if switch A (agency liability) is on, missing either the written advertising authorisation or the sign-off on external material means not a single correction letter may be sent. For list pages whose title is itself laudatory (Best / Top N types), regulated sides make no submission and buy no paid listing — disclosing the payment is not a release from liability; the regulator has specifically answered this question (rule text in the industry edition's appendix (example industry: Dental & Aesthetics Edition, Appendix A)). If you are already listed on one, the strict side asks the publisher to remove you and does not send a correction letter (reasons in → General Edition 5.3 Quick reference by side (2): testimonials and reviews, comparisons, lists, titles, outbound links, FAQ and captions); the light side follows its industry edition's rules and may correct only the one line that is wrong.

If the same error hasn't cleared after two months in a row: escalate for a review, and suspect first that you haven't found the right source — the error may be coming from a third-party source you haven't identified yet.

Make before-and-after screenshots anyone can reproduce

Three rules for anything public: ① Show only the current state — never "I can change it to X" — and no quantity words or timeframe words at all. ② Any price figure AI has made up is shown only as a risk illustration, never written into any external deliverable, and nothing may suggest that anyone can control AI's output. ③ Question 6 (versus a competitor) never goes external.

Entity errors jump the queue

Also not a trigger: an increase or decrease in being listed as a source, or a competitor entering or dropping out — these go to the monthly report, not the queue.

Cost and red lines: jumping the queue draws on that month's content-page hours — it's an open cost, not extra time found from somewhere; it doesn't change the acceptance measure, and even once it's fixed, the only thing you may write is "baseline X errors → this month Y errors", never "I got AI to change its story"; no new content pages during a queue-jump. The clock starts at "discovery": discovery = anyone reading it in any batch of answers.

3.6 The positioning-statement myth, and the exits from this chapter

What you'll do in this section: don't spend 3–4 hours writing one "positioning statement" — put that time into 6 answer sentences for the first intent cluster instead; then check against the checklist at the end of the chapter to decide whether you can move on, or whether you need to jump the queue, push everything back, or do website-side work only.

Why a positioning statement does not work

However solemnly the positioning statement is signed off, it still only exists in a folder. The real, legitimate outlet for that need is the answer sentence: the first sentence under every H2 on a thick page, containing the brand name plus one number or one clear condition, and readable without context — that's a string on the page, and AI can copy it. A positioning statement isn't.

Take the 3–4 hours that would have gone into the positioning statement, and spend it instead writing 6 answer sentences for the first intent cluster — finalised and added to the thick-page brief the same day; use the time you save for the person pages in 3.3 (40 minutes × 3–5 people).

A red line on wording: never say publicly "I'll get AI to describe us this way" — the alternative wording is in → General Edition B.2 Approved wording and scripts. Writing it correctly is your job; whether it gets copied is the machine's.

Three more rules for regulated sides: every public statement must be verifiable, with a source and verified date on each line; personal credentials are not a back door around organisation-level advertising restrictions; on the strict side, price may only be written as a fixed final price — never a range, never "from".

Exits from this chapter

The four exits map to:

• Fix first, then move on: a wrong identity, a merged entity, or any factual error → start jumping the queue the same day (3.5), fix every source you control and send the correction letters, then retest the same question two weeks later; don't start page work on that intent until it's fixed.
• Push everything back: the door layer is blocking retrieval crawlers, or the price is in JS, and it can't be fixed within four weeks → back to → General Edition 2.9 Door-layer troubleshooting and this chapter's acceptance checks; push page and off-site work back as a whole; while the door is shut, however well this chapter is done, it's multiplied by 0.
• The whole lever fails: on a regulated side, you can't get written advertising authorisation → not one piece of outbound material can go out, not even a correction letter; at this point only website-side work is possible (/facts + self-filled profiles), and the whole main line of six-trace alignment and third-party correction is gone.
• Move on: with everything else in place, go to → General Edition 4.1 The full target-picking process and the twelve pre-start questions.

Checklist for this chapter

Common mistakes