What E-E-A-T signals actually move AI citations

E-E-A-T signals that move AI citations are the ones a retrieval system can extract and a reader can check: a named author who matches the visible page, first-hand evidence on that URL, entity facts that match About and schema, original numbers with a method, and third-party mentions on domains the engine already quotes. There is no E-E-A-T scoring API. Google’s quality raters use the framework to grade experiments; operators watch citation logs and fact-match, not a hidden grade.

This spoke sits under the Answer Engine Optimization playbook. I have been SEO certified since 2021. That credential is a checkable claim, not a citation-rate lift I will invent for your deck.

The short answer

• Treat E-E-A-T as a rater and quality framework. Google’s helpful-content documentation says it is not a specific ranking factor.
• Trust is the letter Google calls most important. The others feed it. A pretty author box on a page that lies about price still fails.
• Answer engines cite passages they can lift. Your job is extractable proof, not an “E-E-A-T package” from an agency menu.
• Measure cited / mentioned / absent / hallucinated on a frozen prompt panel. Do not report an E-E-A-T score.
• Off-site corroboration and original research are real layers. They live in the digital PR and original research spokes — not in a badge wall.

What is E-E-A-T, and what is it not?

E-E-A-T is Experience, Expertise, Authoritativeness, and Trustworthiness. It lives in Google’s Search Quality Rater Guidelines. Raters use it to judge whether search experiments look helpful. Google then trains automated systems against that quality bar.

It is not a field in Search Console. It is not a score you can purchase. It is not a ChatGPT setting. Search Central’s wording is the line to quote: while E-E-A-T itself is not a specific ranking factor, Google’s systems use a mix of factors that identify content with those qualities. Raters do not flip rankings on a URL you shipped Tuesday.

Object	What it is	What it is not	What an operator can observe
E-E-A-T	Rater and quality framework	A ranking API, a dashboard metric, a plugin	Self-assessment against public rater concepts
Quality raters	Contracted humans grading experiments	People who “approve” your page	You cannot see their tickets. Do not wait for them.
Ranking systems	Many automated factors at scale	“The E-E-A-T algorithm”	Rank, indexing, snippet eligibility
AI Overviews / AI Mode	Generative units on Search	A separate E-E-A-T product	Cited, linked, or absent in the unit
ChatGPT Search / Perplexity	Retrieval plus a source panel when search ran	Google raters under another name	Sources panel vs memory-only answer
Your log	Screenshots of prompts and cited URLs	An agency “authority grade”	Inclusion and fact-match over time

YMYL — Your Money or Your Life — is the topic class Search Central says gets even more weight when the systems look for strong E-E-A-T: health, money, safety, civic welfare. A cannabis catalog, a medical explainer, and a pricing page are not the same risk as a discography. Do not copy a lifestyle-blog author template onto a YMYL URL and call it done.

Topic class	E-E-A-T bar (rater idea)	What AI systems do with a miss	First artifact
YMYL (health, money, safety, civic)	Highest. Trust vetoes the rest.	Wrong advice is worse than omission	Named reviewer + sources + scope
Commercial offer / pricing	High. Hallucinated price is a sales incident.	Invented packages, cities, retainers	One offer sentence everywhere
First-hand ops / how-to	Experience can outrun credentials	Generic steps, no gotchas	Method + screenshot
Arts / catalog / discography	Accuracy still matters; lower YMYL heat	Wrong release years, fake awards	Canonical facts table
Pure opinion / essay	Who still has to be real	Ghost experts	Honest byline, no fake doctorate

Procedure — use the rater guidelines without treating them as an API:

1. Read Search Central’s helpful-content page (E-E-A-T plus Who / How / Why). That is the operator summary.
2. Skim the current Search Quality Rater Guidelines PDF for Page Quality and YMYL. Google revises it; do not treat a 2022 recap as current.
3. Translate each letter into one screenshotable artifact on your money URLs.
4. Reject vendor language that says “E-E-A-T score,” “E-E-A-T API,” or “E-E-A-T ranking factor.”
5. Put the artifacts in HTML a crawler and a search-enabled chat fetch can see.
6. Measure with the citation log. Never with a rater-imitation spreadsheet.

Operator rule: if a vendor cannot show the artifact (byline HTML, earned URL, method block) and can only show a slide labeled “E-E-A-T,” they are selling a mnemonic.

Which E-E-A-T signals can you actually observe?

You cannot observe a rater’s Page Quality label on your URL. You can observe HTML, entity consistency, earned mentions, and whether an answer engine credits you.

Map each letter to an artifact you can screenshot. If you cannot screenshot it, it is not an operating signal.

Letter	Rater idea (plain)	Artifact you can inspect	Citation-side observation
Experience	First-hand involvement	Original photos, screenshots, n, method, “we ran X”	The answer quotes your method or your number, not a generic how-to
Expertise	Demonstrable skill	Bylines, author URL, credentials that exist, reviewer line on YMYL	The engine names the person or the org correctly
Authoritativeness	Others treat you as a source	Earned mentions, citations, sameAs, press URLs	A third-party page is retrieved for your fact — or your URL is
Trust	Accuracy, honesty, safety	HTTPS, contact, About, corrections, disclosures, fact match	The answer does not invent your price, city, founding year, or offer

Who owns the letter inside a small team — decide before you open twenty tickets:

Letter	Default owner	Artifact they ship	Escalation
Experience	Operator / practitioner who did the work	Method, n, photos, failure notes	If marketing writes it, it is not experience
Expertise	Named author + editor	Bylines, author URL, real credentials	YMYL reviewer if the author is not the expert
Authoritativeness	PR / SEO with the citation log	Earned URLs on already-cited hosts	Legal if Wikipedia / paid disclosure
Trust	Founder or ops	Fact sheet, contact, corrections, HTTPS	Incident channel for “the AI said we…”

Google’s public self-check is Who / How / Why. Translate it into tickets, not vibes.

Google question	Passes when	Fails when	Ticket
Who created this?	Visible byline; author page with checkable bio	“Staff writer,” stock photo, schema-only Person	Author HTML on the money URLs
How was it created?	Method, sample, test notes, or an honest automation disclosure	Date bump; “AI-assisted” footer with no method	Method block next to the claim
Why does this page exist?	Helps a buyer who landed directly	Exists to harvest informational clicks	Kill or merge the doorway

Checklist you can run without a consultant:

• Every money URL has a visible author or organization that matches Person / Organization JSON-LD
• Author page states experience that a stranger could verify (role, years that are true, certifications that exist)
• About, footer, schema, and the first paragraph agree on name, offer, and location
• Claims with numbers sit next to date, population, and method — or they are cut
• Contact and corrections are crawlable, not buried in a PDF
• You have a frozen prompt list, not a mood about “authority”

I will put receipts I can defend on a sales call on an About page: SEO certified since 2021, 500+ automations built, 20,000+ hours on agentic systems, 35,000+ hours saved for clients, hundreds of production sites. I will not put a fabricated “+X% AI citation rate after we added E-E-A-T.” If you cannot say the number out loud, it does not belong in the passage you hope a model lifts.

Which signals move AI citations versus quality theater?

Answer engines retrieve pages, lift a passage, and attach a source when the product shows citations. OpenAI’s ChatGPT Search help is the operator check: inline citations and a Sources panel appear when search ran. No Sources control means the answer came from memory. You cannot “E-E-A-T” your way into a training-weight hallucination the same way you earn a retrieved URL.

Google’s AI features documentation still requires the supporting link to be indexed and snippet-eligible. There is no Overview-specific E-E-A-T markup. A nosnippet accident on the answer block kills the job before any author box matters.

Signal	Moves inclusion (directional)	Theater (looks busy, rarely retrieved)	Why
Answer-first lead plus one table	High	“Ultimate guide” wind-up	Models lift bounded objects
Named author matching HTML	Medium, especially YMYL	Schema Person with no visible byline	Who has to be on the page people retrieve
First-hand method / n	High on how-to and review queries	“In our experience” with no artifact	Experience is evidence, not a tone
Original, dated statistic	High when attributed	Recycled industry “studies”	Credit needs a unique number; see the research spoke
Earned mention on a domain already cited	High as corroboration	Homepage link on a DR-50 lifestyle blog	Retrieval follows language the engine already uses
Entity fact sheet enforced	High as hallucination control	Three founding years across the site	Conflicting facts teach the model to guess
HTTPS, contact, About	Floor	Trust-badge PNG row	Floor is not a differentiator; missing it is a veto
Guest-post networks	Low / negative	“10 authoritativeness links this month”	Wrong category copy becomes the quoted line
Date bump, no substance	Negative vs Google’s own warning	Freshness theater	Search Central lists this as search-engine-first behavior
Identical E-E-A-T footer on 200 URLs	None	Plugin boilerplate	Duplicate chrome is not expertise

Ahrefs’ 75,000-brand analysis put branded web mentions at a 0.664 correlation with AI Overview brand visibility, against 0.218 for referring domains. That is a correlation, not a law, and not “E-E-A-T caused citations.” It is still the cleanest public split I have seen between “we bought a link” and “the model has language about us.”

GEO (Aggarwal et al., KDD 2024) found statistics, quotations, and source citations lifted share of the generated answer on their lab metric. That is visibility inside an answer, not traffic, and not an E-E-A-T multiplier. The method and the hedges live in original research for AI citations. Do not paste “+40% citations from E-E-A-T” into a Q3 plan.

Go / no-go before you fund a “quality” ticket:

• Can a crawler see the artifact in fetched HTML?
• Could a competitor publish the same sentence unchanged? If yes, it is not experience.
• Does the host already appear in your citation log? If no, treat PR as optional SEO.
• Will you retest a frozen prompt 4–6 weeks after indexation? If no, you will not know if it moved.
• Are you about to invent a person, a credential, or a statistic? Stop. That is a trust incident, not a gap.

Theater fails in a specific way: the page looks “authoritative” to a marketer and still has nothing a model can quote that competitors do not also have.

How does Experience show up in an answer engine?

Experience is first-hand involvement. Google’s helpful-content page uses product use, visiting a place, and first-hand expertise as examples. Raters are looking for evidence the writer did the thing. Retrieval systems are looking for a sentence that is not a paraphrase of the ten pages above you.

A music site that lists tour dates the artist actually played is experience. A paraphrased “best alt-rock openers” roundup is not. A shop post that shows the pack you ship is experience. A wellness rewrite of a PubMed abstract is not — and on YMYL-adjacent topics it is a trust problem.

Query class	Experience that can be cited	Fake experience	What to put on the page
Product / review	Test protocol, n, photos you took	“We tested everything” with stock shots	Table of tests + date
How-to / ops	Screenshots of the live UI, failure notes	Generic steps copied from docs	Numbered procedure with the gotcha
Local / services	Job photos, license, service area you actually cover	City-page mill	NAP + one proof of work
YMYL-adjacent	Practitioner credentials + what you personally observed	Medical tone from a marketing blog	Scope the claim; cite a primary source
Brand story	Founding facts that match filings and About	Mythology and “trusted by leaders”	One paragraph a model cannot invent

Procedure for one URL:

1. Write the lead answer in 40–80 words without the word “experience.”
2. Add the first-hand object: photo, screenshot, table, or n.
3. Label how you know: date, tool, sample, or “this is the workflow we run.”
4. Cut any sentence a competitor could publish unchanged.
5. Re-fetch the URL. Confirm the object is in the HTML, not a client-only widget.
6. Ask the frozen prompt. Log whether the answer used your object, a competitor’s, or a blend.

Checklist:

• There is an artifact that would be expensive to fake this week
• The artifact is next to the claim, not in a footer
• The page does not hide the method behind a lead form
• Failure modes are named (what broke, what you did instead)
• You would defend the artifact on a call without flinching

“I have 20,000+ hours on agentic systems” is experience only if the page then shows a system, not if the bio is the whole proof. Bravery is not a restore strategy, and a bio is not a method.

How does Expertise show up without a fake bio?

Expertise is demonstrable skill. Search Central asks whether the content is written or reviewed by someone who knows the topic, and whether a researcher of the site would walk away thinking it is recognized. That is not a request for a novel-length About page. It is a request that Who is true.

Fake bios are the most common E-E-A-T self-own I still see: “Dr. Jane Hale, 15 years,” stock headshot, no LinkedIn, no byline history, schema Person stuffed into 90 posts. The rater framework treats that as a trust hit. Retrieval will still quote the page if the passage is convenient — and then your invented doctor becomes the named source. That is worse than being omitted.

Expert pattern	Do this	Do not do this	Check
Real operator	Legal name, role, one checkable credential, author URL	Collective “our team of experts”	Google the name. Does a real trail exist?
Reviewer on YMYL	Named reviewer + what they reviewed	Disclaimer instead of a reviewer	Reviewer line in HTML
Organization-only	Clear org author on policy/docs pages	Person schema for a ghost	Organization matches the byline
Guest expert	Disclosure of relationship	Paying for a quote and hiding it	FTC-ish honesty; see the PR spoke
Credentials	Ones that exist (dates, issuers)	“Certified visionary”	Issuer page still live?

I use credentials that survive a lookup: SEO certified since 2021, all Make.com AI automation certifications, direct collaborations with the n8n team. I do not invent a medical byline for a marketing URL. If the page is not in my lane, I do not play expert on it.

Procedure:

1. Inventory bylines on the 20 URLs in the frozen panel.
2. Delete schema Person nodes that do not match visible HTML.
3. Build or fix one author page per real writer. One. Not forty placeholders.
4. Match sameAs only to profiles you control and that agree on the facts.
5. On YMYL, add a reviewer only if that person actually reviewed the claims.
6. Re-run three prompts that previously named a competitor’s expert. Log whether your name appears. Do not call a miss a “score.”

Checklist before you publish an expert chrome pass:

• Every Person in JSON-LD has a matching visible byline
• Author URL 200s and states only facts you would repeat on a call
• sameAs targets agree on name, role, and employer
• YMYL pages name a reviewer who actually read the claims
• No leftover “staff writer” or stock-photo doctors
• Three trap prompts re-run after indexation; results logged, not graded

If you have one expert and eighty thin posts, the fix is fewer posts, not more author boxes.

How does Authoritativeness show up off your domain?

Authoritativeness is other people treating you as a source. Raters look for recognition. Retrieval looks for corroborating language on URLs it already trusts.

Your homepage saying you are the leading studio is not authoritativeness. A trade publication repeating your one-sentence category line can be. A Wikipedia page you secretly wrote is a policy problem, not a win. Digital PR for citations is the operating spoke: media list equals citation log, lock the fact sheet first, measure inclusion.

Off-site object	Useful when	Useless when	Citation watch
Earned article with a quote-ready line	Domain already appears in your answer log	DR-only placement, no topical sentence	Does the engine start quoting that line?
Unlinked brand mention	Language about you exists	Mention is a tag soup	Ahrefs-style mention correlation is directional
sameAs to Wikipedia / Wikidata / LinkedIn	Those pages are accurate	You stuffed five socials that disagree	Fact-match across the graph
Guest post	You control the sentence	Network farm, wrong category	Read the live URL. Would you want it retrieved?
Awards / “as seen in”	The appearance is real and dated	Logo wall of unpaid or fake marks	Click every logo. 404s fail Trust too
Directory citations	NAP consistent for local	Spam citations for a national brand	Consistency over volume

Ahrefs’ mention-vs-link split belongs here, not in a backlink report. If the engine never retrieves the host, the placement is optional SEO PR.

Procedure:

1. Export the last 30 days of cited domains from the prompt log.
2. Pitch only hosts already on that list, or hosts that consistently appear for the category even if they have not named you yet.
3. Give editors a 40–80 word passage and a criteria table. Adjectives get rewritten.
4. Lock founding year, offer, and category before anyone files the pitch.
5. When the piece goes live, add the URL to the retest set. Wait 4–6 weeks. Inclusion is the KPI, not Domain Rating.

Checklist before a pitch leaves the building:

• Fact sheet signed off (founding year, offer, category, location)
• Target host appeared in the citation log or category answers in the last 90 days
• Passage is 40–80 words plus a table the editor can keep
• Relationship / payment disclosure path is honest
• Retest date sits 4–6 weeks out on the same prompts
• Someone is assigned to read the live URL the day it ships

Do not buy a “10 authoritative links” package and map it to this letter. Mapping is how theater gets a budget.

Why does Trust gate the other three letters?

Search Central is explicit: of the E-E-A-T aspects, trust is most important. The others contribute to trust. Content does not have to demonstrate every letter. Low trust still vetoes the rest. A first-hand story on a site that spoofs contact info does not become citeable because the photos are real.

For AI answers, trust shows up as fact consistency. Models blend sources. If About says Northern Michigan, a guest post says Los Angeles, and schema says New York, the answer will pick one and sound sure. You will spend the next quarter correcting a machine.

Trust artifact	Pass	Fail	AI failure mode
Entity facts	One sheet; About, footer, schema, PR agree	Three founding years	Hallucinated origin story
Contact	Crawlable, real, monitored	Form that 500s; no legal name	“Is this even a company?”
HTTPS / malware	Clean	Mixed content, injected spam	Eligibility floor
Corrections	Dated changelog on claims that moved	Silent rewrites, date bumps	Engine keeps the old number
Disclosures	Affiliate, AI-assist, sponsored — labeled	Native ads dressed as editorial	Policy plus trust
Reviews / complaints	You respond in public where it is real	Fake review inventory	YMYL especially
Snippet eligibility	Answer block indexable	nosnippet, noindex accidents	You cannot be the supporting link

Google’s search-engine-first list is a trust list in disguise: mass automation for rankings, rewriting other people with no added value, promising a release date that does not exist, changing dates without changing content. Those behaviors are documented on the same helpful-content page as E-E-A-T. They are not a separate “spam lane” you can ignore while you “work on authority.”

Checklist:

• One entity fact sheet: legal name, brand name, offer sentence, location, founding year, credentials
• Diff About vs schema vs the last three PR clips
• URL Inspection on the five money URLs: fetched HTML contains the answer
• No max-snippet:0 on the answer block
• Correction policy exists as a page, not a Slack intention
• Someone owns inbound “the AI said we…” tickets

Trust is the cheapest letter to wreck and the slowest to repair. Protect it like a production credential, because that is what it is.

What do Google, ChatGPT Search, and Perplexity actually extract?

They do not extract an E-E-A-T vector you can dump from an API. They extract text, structure, and (when search is on) URLs.

Hedge: product behavior changes. As of the August 2026 docs I am citing, Google still says no extra Overview markup; ChatGPT Search still distinguishes memory from search via the Sources panel; Perplexity still shows numbered sources. None of them publish “we score E-E-A-T.” Overlap with the rater framework is practical: clear Who, first-hand How, accurate facts, corroboration.

Surface	What you can see	What you cannot see	Operator move
Google AI Overviews / AI Mode	Cited links; Search Console generative AI impressions if rolled out to the property	Internal quality grade	Indexed, snippet-eligible, extractable lead + table
Classic Search	Rank, sitelinks, People Also Ask	Rater labels	Same quality bar; still not an E-E-A-T score
ChatGPT Search	Sources panel when search ran	Weights, memory vs browse mix when the panel is absent	Put the citeable sentence in HTML; retest with search on
Perplexity	Numbered citations	Why source 3 beat source 1	Be a source that answers in a bounded block
Gemini / other chat UIs	Variable citation UX	Almost everything	Do not chase every wrapper; freeze two products plus Google

Extraction preferences that show up in practice — directional, not a vendor SLA:

1. Short, declarative lead. The first 60 words should stand alone.
2. Tables, numbered procedures, bounded checklists. Prose mush ties.
3. Attributed numbers. “We surveyed 40 customers in July 2026” beats “most clients see results.”
4. Consistent named entities. Org, person, product, place spelled the same way everywhere.
5. Third-party repeats. One fact on three reputable URLs beats one fact on your blog and nowhere else.

Procedure for a single prompt:

1. Run the prompt in a logged-out or documented session (product, date, logged-in state).
2. Screenshot the answer and the source list.
3. Classify: cited with link, mentioned without link, absent, or hallucinated fact.
4. Diff the quoted sentence against your HTML. If they quoted a competitor’s table, that is the ticket.
5. If they quoted a third party about you, read that third party. Fix the fact sheet if it is wrong.
6. Do not average this into an “E-E-A-T lift %.”

Checklist per product you bother to log:

• Search / sources panel actually ran (or you labeled the run as memory-only)
• Quoted sentence exists in your HTML, a third-party URL, or neither
• Entity fields (name, offer, place, price) match the fact sheet
• Supporting link is indexable if the product is Google
• You captured date, product, and login state so next week is comparable

If the engine cites you for a sentence you did not write, you have a corroboration problem, not a missing badge.

How do you measure whether the signals are working?

You measure inclusion and accuracy on a frozen set. You do not measure E-E-A-T.

Build a panel of 20–40 prompts that match how buyers ask. Include brand, category, comparison, and one trap prompt that used to hallucinate a fact. Hold the wording for 90 days. Tools that estimate “AI traffic” are optional color. They are not the scoreboard.

KPI	Source	Cadence	Kill / act criterion
Citation rate	Manual log: cited ÷ Overview-or-search prompts	Weekly	0% for a quarter on money prompts → reinvent the page, do not buy bios
Mention without link	Same log	Weekly	Language exists; pitch the missing URL via PR
Hallucinated fact	Same log	Weekly	Same-day fact-sheet + source correction
Fact-match score	Answer vs entity sheet (binary per field)	Weekly	Any fail on price, location, offer is P0
Generative AI impressions	Search Console report if the property has it	Weekly	Impressions without citations → extractability job
Brand queries	Search Console	Weekly	Collapse after you “pruned the blog” is a demand hole
Protectable conversions	Analytics / CRM	Weekly	The number that pays rent
Referring DR	Ahrefs / equivalent	Monthly	Report as SEO, never as AEO success

Procedure:

1. Freeze the prompt list in a sheet. No drive-by additions mid-test except a dated appendix.
2. Capture product, date, device class, and login state.
3. Two reviewers classify cited / mentioned / absent / hallucinated. Disagreements go to a third screenshot, not a debate.
4. Ship one change cluster at a time (authors, or research, or PR). Mixing them makes attribution folklore.
5. Retest 4–6 weeks after indexation, not 48 hours after the CMS save.
6. Report counts and examples. Never a made-up lift.

Measurement checklist (print this above the sheet):

• Prompt wording frozen; appendix dated if you add one
• Two-person classification on cited / mentioned / absent / hallucinated
• One change cluster per window (authors or research or PR)
• Retest clock starts at indexation / live PR URL, not at CMS save
• Adjacent studies quoted with name and date, never as your lift
• Exec slide has counts and screenshots — no “E-E-A-T score”

I will not invent a citation-rate improvement from adding E-E-A-T chrome. Public studies that look adjacent (Ahrefs mentions, GEO evidence-adding edits) are not your property’s result. Quote them with method and date, then show your panel.

What breaks when you buy an E-E-A-T package?

The failure mode is an eight-week “authority sprint” that ships chrome and poisons the entity graph.

Typical package: forty author pages for writers who never wrote, a logo wall, a dozen guest posts that call you an AI agency when you are a studio, schema stuffed until the validator wheezes, a date bump across the blog. Leadership hears “we did E-E-A-T.” The frozen panel still cites competitors. Worse: ChatGPT Search now says you were founded in 2014 because a guest post needed a narrative. You will spend real money undoing a sentence you paid to publish.

Failure	Cost	What it looks like in the log	What you do instead
Ghost authors	Trust + legal risk if YMYL	Engine names a person who does not exist	One real byline
Wrong-category PR	Hallucination fuel	Answers describe the wrong offer	Fact sheet before pitch
Badge wall	Design time, zero retrieval	No change in sources	Delete; link only real appearances
Mass thin experts	Helpful-content pattern	Pages look like a content network	Fewer URLs, deeper proof
Schema/HTML mismatch	Inconsistent Who	Bylines disagree with JSON-LD	HTML wins; fix schema to match
Date bumps	Explicit Google warning	Fresh dates, stale claims	Touch dates only when substance changes
Buying DR as authoritativeness	Budget + junk hosts	Cited domains never match the purchases	Citation-log media list

I have watched this pattern on sites that already had the unglamorous receipts — hundreds of shipped pages, years of SEO work — and still got sold a mnemonic. The package did not fail because E-E-A-T is “fake.” It failed because the vendor treated a rater framework like a checkout SKU.

Recovery procedure:

1. Stop new guest posts and author-plugin spam the day you notice the hallucination.
2. Publish a corrections URL. Date it.
3. Align About, schema, and the three live clips that are wrong. Email editors if they will still edit.
4. noindex or merge ghost author pages. Do not leave 40 doors open.
5. Re-run the trap prompts weekly until the invented fact dies.
6. Only then resume PR, and only to hosts on the citation log.

An invented expert is not a missing ingredient. It is a contaminated source.

Which pages should get the signals first?

Not the whole blog. Overviews and chat search punish undifferentiated libraries. They do not reward 80 identical author boxes.

Start where a citation pays rent or where a hallucination costs a sale.

Page type	Why it is first	Signal to add	Do not
About / Author	Entity source of truth	Fact sheet, sameAs, real bio	Mythology
Contact / legal	Trust floor	Real identity, policy, corrections	Chat widget only
Offer / pricing / audit	Money + hallucination magnet	Exact offer sentence, dated if prices move	Vague “packages start at”
3–5 informational URLs already in the panel	Already retrieved or close	Lead answer, table, byline, method	Twenty new outlines
Case / work pages	Experience evidence	Named work you can stand behind, no invented metrics	Fake outcomes
YMYL explainers	Heightened bar	Reviewer + sources + scope	Medical voice from marketing
Thin glossary	Usually later / merge	Citeable definition or kill	Author chrome on stubs

Portfolio names I will mention without attaching invented numbers: Fuller, Oliver Malcolm, Arkayla, Dog Park, Divine Toke, AllCity HVAC. A work page that states what shipped is experience. A work page that invents a traffic lift is a trust problem.

Prioritization procedure:

1. Take the frozen panel. List landing URLs the engines already cite (even if not you).
2. Map each prompt to one owned URL that should win, or to “needs a new URL.”
3. Score each owned URL: Who visible? How evidenced? Facts match sheet? Extractable object? Snippet-eligible?
4. Fix the highest-impression, lowest-score URLs first — usually About + two explainers + the offer page.
5. Merge stubs that cannot earn a unique object in one rewrite.
6. Open no more than five content tickets in the first two weeks.

Checklist for the first five tickets:

• About and offer pages quote the same fact sheet
• Each ticketed explainer has a lead answer and one table
• Who is visible without executing JavaScript theater
• Work/case pages use only outcomes you can defend (no invented traffic lifts)
• Stubs without a unique object are merged, not decorated with author boxes
• Cap of five open content tickets for two weeks

If the About page disagrees with the offer page, you do not have an E-E-A-T problem on the blog. You have an entity problem at the root.

What should you skip if you only have a week — and what belongs in ninety days?

A week is for observability and the vetoes. Ninety days is for corroboration and original evidence. Mixing them is how you get a redesign and no baseline.

One week (do these; skip the rest)

Day	Action	Done looks like
1	Freeze 20–40 prompts; screenshot current answers	Sheet with cited / mentioned / absent / hallucinated
2	Write the entity fact sheet	One page the whole company can quote
3	Diff About, footer, schema, top PR clips	Ticket list of contradictions
4	Bylines + author/org HTML on five money URLs	Fetch as Google; Who is visible
5	One extractable table on the highest-value explainer	Lead answer + table in HTML
6	Kill nosnippet / ghost authors / fake badges	Eligibility floor restored
7	Retest the five prompts most likely to hallucinate	Same-week fact-match, not a lift claim

Skip in week one: Wikipedia campaigns, national PR retainers, buying DR, rewriting the archive, author pages for people who do not write, anything a vendor branded “E-E-A-T package.”

Ninety days

Window	Ship	Proof
Days 8–30	Remaining money-URL extractability; correction policy live	Panel still frozen; weekly log
Days 31–60	One original number with method, or a stop decision that you will not fake a study	Research spoke rules; n and date in HTML
Days 31–75	Digital PR only to domains already in the citation log	Final URLs, not impression decks
Days 60–90	Kill/merge list; exec one-pager with inclusion counts	No E-E-A-T score on the slide

Weekly ops after week one:

• Ten-prompt subsample logged (full panel monthly)
• Hallucinations assigned an owner the day they appear
• One shipped HTML change or an explicit defer
• PR and research queued only if the fact sheet is still clean

Blockers that make the letters premature:

Blocker	Why the letters wait	Do this instead
Money URLs noindex / blocked	Nothing to retrieve	Fix robots, noindex, and Inspection
nosnippet on the answer	Google cannot use you as a supporting link	Remove the accidental directive
No one-sentence offer	Every PR clip will disagree	Fact sheet before chrome
No frozen prompts	You will argue from vibes	Twenty prompts, one sheet
YMYL with no expert	Trust veto	Scope the claim or get a real reviewer
Vendor wants a 40-author plugin	Ghost-expert failure mode	One real byline

When this is not worth doing yet: the money URLs are noindex, the business cannot state a one-sentence offer, or nobody will freeze a prompt list. Fix eligibility and honesty first. A rater mnemonic will not rescue an unindexed page.

A visibility audit is the right hammer when entity conflicts, YMYL reviewer gaps, and citation logs disagree and the argument has gone political. DIY the week-one table if you have Search Console and a writer who will put their name on the work. Lane map: /visibility.

FAQ

What E-E-A-T signals actually move AI citations?

The ones a system can extract and a stranger can check: named authors that match the HTML, first-hand method or n on the page, entity facts that agree across About and schema, original attributed numbers, and third-party mentions on domains the engine already quotes. Google describes E-E-A-T as a rater and quality framework, not a specific ranking factor, and not a score. Fake bios, badge walls, and DR packages are theater.

How do I measure whether E-E-A-T signals are moving AI citations?

Freeze 20–40 prompts and log cited, mentioned, absent, and hallucinated, plus fact-match against an entity sheet. Use Search Console’s generative AI impression report if the property has it. Do not invent an E-E-A-T score. Retest 4–6 weeks after indexation, and ship one change cluster at a time so you can see what actually moved.

What usually fails first when teams try this?

Ghost authors and guest posts that mint the wrong category line. The frozen panel does not move, and a chat product starts repeating the invented founding year or offer. Schema that does not match visible bylines is a close second. Fix HTML and the fact sheet before you buy corroboration.

How long does this take to show results?

Eligibility and contradiction fixes can show up in the next crawl and the next search-enabled chat run — sometimes days, often a few weeks. Earned mentions and original research are slower: plan a 4–6 week retest after the URL is live. There is no citation SLA, and I will not quote a fabricated lift curve.

What should I skip if I only have a week?

Skip Wikipedia campaigns, national PR, DR packages, archive rewrites, and author pages for people who do not write. Spend the week on a frozen prompt log, an entity fact sheet, bylines on five money URLs, one extractable table, and killing nosnippet plus fake badges.

When is this not worth doing yet?

When money URLs are not indexable, when the company cannot agree on a one-sentence offer, or when nobody will keep a prompt log. E-E-A-T work on top of that is chrome. Get fetched HTML and honest facts first; then the letters have something to attach to.

CTA

If you have been sold an E-E-A-T score, throw the slide out — lock the facts, put Who on the page, and log whether the engines actually cite you.

Lane: /visibility · Book a visibility audit.