AEO audit Runs on your machine, on a local model

How do I optimize this
page for AI search?

Point Moose at a URL and he reads it the way a crawler does, not the way you do. What comes back is ordered the way it has to be fixed: can an engine reach the page, can it lift an answer out of it, does it have any reason to trust it, and only then, is the copy good.

Audit a page See what it hands back
Every recommendation arrives with the copy already written. No word counts, no schema advice, no promises about citations.
Chat Workflows
Meridian Payrollhttps://meridianpayroll.com
New chat
Home Inbox Context Visibility Workflows Chats Library Connections
Recents
Run an AEO auditChat · now multi state payroll complianceAEO audit · 6d
D DanaBYOK Free
Run an AEO audit New chat Connect your CMSW
How do I optimize this page for AI search? meridianpayroll.com/multi-state-payroll-compliance

I can audit that page. Confirm the source below and add the query you want it to win, if you have one in mind. If you leave it blank I will infer one from the page and audit against that.

Confirm the source and target query
Target prompt or query
multi state payroll compliance
Optional, but helpful for stronger evidence collection.
Source content
URL Paste Upload
https://meridianpayroll.com/multi-state-payroll-compliance
Run AEO audit
Audit complete 2:34
Fetching the raw HTML Came back as a JavaScript app shell. 2.1k characters of readable text. Retrying
Retrying through a managed browser Rendered 18.4k characters. Static coverage 11%, and no major AI crawler except Googlebot runs JavaScript. Critical
Screenshotting with JavaScript disabled Layout, scan path and whether the answer is visually obvious judged on that view.
Reading robots.txt against 12 crawler tokens 10 of 12 allowed. Claude-SearchBot is disallowed, which is a retrieval crawler. Critical
Extracting the signals Title, meta description, canonical, 14 headings, 0 date signals in schema or visible text. Notice
Running answer readiness Atomicity, ambiguity, entity consistency, direct answer placement. 6 paragraphs flagged. Notice
Probing internal links 31 sampled, 2 broken, 1 chain of four redirects. AI crawlers do not retry a failed fetch. Notice
AEO audit Baseline run
multi state payroll compliance
73/100 Good optimization. Solid foundation with some areas to strengthen. Two findings outrank the score itself: the page is invisible to Claude's retrieval crawler, and there is almost nothing in the static HTML for an engine that does not run JavaScript.
AI crawler access 10 of 12 allowed
Retrieval ✓ OAI-SearchBot ✗ Claude-SearchBot ✓ PerplexityBot ✓ Bingbot ✓ Googlebot
Live fetch ✓ ChatGPT-User ✓ Claude-User ✓ Perplexity-User
Training only ✗ GPTBot ✓ ClaudeBot ✓ Google-Extended ✓ CCBot

GPTBot is training only, so blocking it does nothing to your ChatGPT search visibility. Claude-SearchBot is retrieval, and blocking that one removes the page from Claude answers entirely.

Scored categories
Structure 12% 88/100
Semantic clarity 12% 84/100
Comprehensiveness 15% 81/100
Extractability 12% 71/100
Citation potential 12% 66/100
Trust signals 15% 62/100
Freshness 15% 54/100
Answer readiness 12% 49/100

Fix the two retrieval findings first. Then this, which is the highest-value content change on the page.

Evidence-backed Answer readiness Replace · hero paragraph
The page does not answer the question until word 640

Query intent reads as commercial-informational, so a dictionary opening is not required. A one-paragraph direct answer is. There is no definition, summary or step list in the first 15% of the page.

Current text on the page

Payroll has never been more complicated. Between remote work, new state filing rules and a compliance landscape that shifts every quarter, growing teams are under more pressure than ever to get it right.

Publish this instead

Multi-state payroll compliance means registering as an employer in every state where an employee lives, withholding that state's income tax, and paying its unemployment insurance. If your team works across state lines, you owe filings in each of those states, not only the one your office sits in.

Stage to WordPress Copy Anchor verified verbatim · QA passed
Export as PDF Save audit Show the other 11 issues
Share a URL, a draft document, or run an AEO audit…
Local · Gemma 4 Send
Hi, Moose 0.3.286 Beta Need help?
A recreation of the real thing, sped up. A full audit takes two to four minutes.
Moose, a dog, supervising a desk

Chief desk supervisor. Meridian Payroll is invented for this page. The findings, the wording of them, and the order they arrive in are all real.

The order that matters

Optimizing for AI search is four jobs, and the order is not negotiable

Most advice starts at the bottom of this list, because rewriting a heading is easy and reading robots.txt is not. Nothing in tiers two to four does anything if tier one fails. That is the order the audit enforces, and the order the recommendations come back in.

01
First
Retrieval access

Can an engine fetch this page, read it without running JavaScript, and follow the links out of it? These are pass or fail, and a failure here means nothing else on this list can help you.

robots.txt, 12 tokens, judged separately Server-rendered text, not an app shell Broken links and redirect chains
Proven

Without these an engine cannot cite the page at all. This is the only tier Moose states firmly.

02
Second
Answer extractability

An engine needs one passage it can lift out and stand behind without the rest of the page. One claim per paragraph, no orphan pronouns, one name per thing, and the answer somewhere near the top.

Atomicity Ambiguity Entity consistency Direct answer placement
Evidence-backed

Supported by large-scale studies of citation behaviour. Strong, not proven for your specific page.

03
Third
Trust and freshness signals

Whether a reader, or an engine, can tell when this was written and where the claims came from. Real dates in schema and visible on the page. Named sources with links out.

datePublished and dateModified Visible byline dates Attribution and external references
Evidence-backed

Engines hesitate on pages they cannot date. Moose will not tell you to add a date it already found.

04
Last
Polish

Heading wording, paragraph rhythm, coverage of the adjacent question. Worth doing and cheap to do, which is exactly why it should not be the week you spend before fixing tier one.

Heading outline depth Coverage breadth Paragraph shape
Best practice

Low cost and sensible, not proven to increase AI citations on its own. Framed that way in the output.

The scorecard

Eight weighted categories, and what each one looks at

The score is one number over eight categories. Every one of them is something you can go and look at on the page yourself.

Freshness 15%

Real date signals, both in schema and visible on the page, and how old they are. A page with no date at all is a page an engine has to guess about.

Comprehensiveness 15%

Coverage breadth and heading outline depth against the query. Never a word count, because there is no minimum word count for AI visibility.

Trust signals 15%

Whether the claims on the page are attributed and sourced. Named sources, links out to them, and evidence a reader could follow.

Structure 12%

Headings, title, paragraph shape. The scaffolding that tells an engine which part of the page is the answer to which question.

Semantic clarity 12%

How readable the meaning is without the surrounding context. Repeated named entities, consistent terminology, plain sentence construction.

Extractability 12%

How easily a self-contained chunk can be pulled out. Lists, tables, short definitional passages, sections that finish their own thought.

Citation potential 12%

Whether there is anything here worth pointing at: original figures, named sources, specific claims rather than restated consensus.

Answer readiness 12%

Four sub-checks on whether a single passage answers the question when read alone. The part of the audit no classic SEO tool runs.

The AI-specific part

Answer readiness: the four checks a classic SEO tool does not run

An answer engine does not read your page. It looks for a passage it can lift out and stand behind on its own. These four checks are about whether any passage on your page survives being read alone, and each one hands back the paragraph that failed.

Atomicity 30% of the sub-score

Flags paragraphs carrying too many claims at once. Four or more sentences, or two-plus sentences stacked with three or more conjunctions or four or more commas. An engine cannot lift a clean answer out of a paragraph that is doing five jobs.

What it hands back 4 claims, 6 commas

Reciprocity agreements between states can reduce the number of returns you file, and while they are common in the Midwest, they do not cover unemployment insurance, which each state administers separately, and they do not apply to local taxes in cities like Philadelphia, so one remote employee may still create three filing obligations.

Ambiguity 25%

Flags paragraphs leaning so hard on pronouns that the subject disappears when the paragraph is read on its own. Which is exactly how an answer engine reads it.

What it hands back Subject unresolvable

It applies to them as soon as they cross the threshold, and once that happens they have to register with it before the next quarter closes.

Entity consistency 25%

Catches a page flip-flopping between a full name and its acronym, and catches topic drift where the dominant entity changes from paragraph to paragraph.

What it hands back 3 names, 1 entity

State Unemployment Insurance in the opening section, SUI from the third heading onward, SUTA in the FAQ. Dominant entity also changes at paragraph nine, from payroll tax to worker classification.

Direct answer placement 20%

Checks whether the first 15% of the page contains a definition pattern, a summary or TLDR, or a short step list, and how many words a reader wades through before hitting it. Query intent is read first, so a commercial page is not punished for skipping a dictionary opening.

What it hands back 0 of 3 patterns, 640 words

Intent read as commercial-informational. The first definition pattern appears 640 words in, after two customer stories. Nothing in the opening 15% answers the query.

How much to trust it

Every recommendation says how confident it is allowed to be

Nobody knows exactly how answer engines pick a citation, so a tool that speaks with one voice about everything is bluffing somewhere. Each item carries a tier, and the tier decides the wording it is allowed to use.

Proven

A retrieval precondition. Without this, AI engines cannot cite the page at all. Stated as fact, because it is testable and it either passes or it does not.

Findings in this tier
A blocked retrieval crawler token. An app shell with no server-rendered text. Broken internal links and long redirect chains.
Evidence-backed

Supported by large-scale studies of AI citation behaviour. Recommended with confidence, but framed as strengthening quotability rather than guaranteeing a citation.

Findings in this tier
A direct answer near the top. Real date signals. Attributed claims. One claim per paragraph. Consistent entity naming.
Best practice

Low cost and sensible, but not proven to increase AI citations on its own. Offered last, and labelled so you can decide it is not worth your afternoon.

Findings in this tier
Heading wording. Paragraph rhythm. Section ordering. Adding a visible question and answer block.
Your own data, in the audit

It uses what you already know about the page

Moose reads your Search Console numbers for the URL and your own AI visibility history for it, then tunes the recommendations to the page in front of him. A page that already performs gets protected. A page nobody has found gets pushed.

Search Console
How the page performs in search today

Connect it once and every recommendation arrives with a risk level attached.

The page already ranks

Moose protects what is working. Support blocks and clearer answer extraction come first, ahead of touching the title, the H1 or the core framing.

The page has headroom

Bigger swings are on the table, including coverage for the adjacent queries the page nearly answers already.

Your AI visibility history
Whether engines have cited this URL

Moose already tracks which of your URLs get cited. He looks back 90 days on this one before he suggests anything.

Cited today The passage engines quote is left alone. Everything else is fair game.
Cited before Moose looks at what changed on the page and what the competitor now being cited has.
Brand named You got mentioned but a different page of yours got cited. This page becomes the better landing spot.
Never cited Retrieval access and one liftable answer come first, before any polish.
What you get on your plan

Every tier runs the full diagnostic. Paid tiers add the outside world.

The crawler check, the JavaScript content gap, link health and all four answer readiness diagnostics run on your own machine on every tier. A key adds a second opinion on the copy. A managed plan brings in live research, competitor grounding and automation.

Free preview
Local model, on your machine

The audit runs entirely on a local Gemma 4 model. There is no managed path, and no monthly cap either, because it is your hardware and your electricity.

Full page fetch, including the managed browser fallback for JavaScript-heavy or bot-protected pages
All heuristic scoring: freshness, structure, extractability, comprehensiveness, semantic clarity, citation potential
All four answer readiness diagnostics
robots.txt across 12 tokens, the JavaScript content gap, and link health probes
Issue catalogue, baseline and verification re-audits, PDF export, save to Library
Not on this tier
Live research. No Google results, no AI Overview, no ChatGPT Search snapshot, no competitor crawls.
The visual pass. The JavaScript-disabled screenshot is skipped, so layout and scan path are not assessed.
The Moose Brain QA judge, so recommendations are checked by the deterministic rules only.
Workflow-triggered audits.
BYOK
Your own OpenRouter key

Audits run through your key, or a local model if that is what you have selected. Bringing a key mostly buys you the second opinion on the recommendations.

Everything in the free preview
The Moose Brain QA judge runs on your key: shape match, verbatim anchor, duplicate coverage, brand voice
The screenshot pass works with a vision-capable model on OpenRouter
Recommendations planned with the full contract, not one issue at a time
Still not included
Live SERP research. That stays a paid managed upgrade.
The QA judge, if you are on a local-only model. The screenshot pass drops there too.
Workflows, on BYOK Free.
Paid managed
Live research and automation

The audit stops being page-only. It reads the competitive picture around the page, and it can run itself.

Live research inside the audit: a search seed, the AI Overview, and a ChatGPT Search snapshot
Up to three competitor pages and two citation pages crawled, feeding competitor grounding into the scoring
The managed Moose Brain passes: edit intent, surface semantics, anchor resolution, improvement judge
Workflows that run an audit automatically when a visibility metric drops
Worth knowing
The highest-evidence findings are the same ones the free tier gets. This buys context, not the fundamentals.
!

On the free tier the audit needs a local model, and if none is loaded it stops instead of guessing: “Open Settings > Local AI model to download and load a local Gemma 4 model, then try the audit again.” Direct Gemini keys are also rejected for the audit passes. OpenRouter or local.

Common questions

Does schema markup help me get cited by AI?

The best available evidence says no, and Moose will never recommend schema.org or JSON-LD as a way to improve AI visibility. Answer engines read visible page text. Schema still has uses in classic search, so this is not advice to remove it. It is a statement that adding it is not an AEO lever. A visible question and answer section on the page is fine to recommend, because a reader can see it.

What about llms.txt or an AI sitemap?

Same answer. Machine-readable files aimed at answer engines have no evidence behind them for citations, so they are on the refusal list. If a tool tells you to publish one and calls it an AI visibility fix, ask what evidence that is based on.

How much does an audit cost me?

On the free tier, nothing beyond your own electricity. It runs on a local Gemma 4 model on your machine, and there is no monthly cap. It does need a local model downloaded and loaded first, and if there is not one it tells you that instead of guessing at an answer.

Will fixing these things get me cited?

Moose is not allowed to promise that, and neither will we. The retrieval preconditions are different: if a retrieval crawler is blocked or the page has no server-rendered text, engines cannot cite you, and fixing that is stated as fact. Everything after that is framed as strengthening quotability and reader trust, because the direct effect on citations is not something anyone can currently prove.

My page is a JavaScript app. Is that fatal?

It is the finding that outranks almost everything else. No major AI crawler except Googlebot executes JavaScript, so if your readable text only exists after rendering, most engines see an empty page. The audit reports the actual numbers: static characters, rendered characters, and static coverage as a percentage.

Do I have to apply the changes myself?

You can. Every recommendation gives you the current text and the replacement copy, so it is a copy and paste. If WordPress is connected you can stage the approved changes and publish them from the audit, and nothing goes live without you approving it.

More things to ask Moose

All use cases
Runs on your machine

Start with the page you
most want cited.

Paste the URL into a chat. You get the crawler check, the content gap, link health, answer readiness, and copy you can publish, in the order it has to be done.

Create the account here, then install the desktop app.