Run 2026-07-30. Five reasoning models, five different labs, each given the full markup of the five money pages, the text of seven supporting pages, the 40-page inventory, and the real business context. Framed as hostile: find what is wrong and what is missing.

Panel: Claude Opus 5 (Anthropic), Gemini 3.1 Pro (Google), GPT-5.5 (OpenAI), Grok 4.5 (xAI), GLM-5.2 (Z-AI). All five returned. 258k tokens in, 55k out, roughly $1.20.

Backlinks and citations were banned from the prompt, as were the items the 2026-07-25 audit already covered and anything requiring invented proof.

---

Before anything else: two findings I killed

I verified every checkable claim against the live site rather than passing the panel's output through. Most held up. Two did not, and both were my fault rather than the model's.

Gemini said the homepage has zero JSON-LD structured data. False. The homepage carries two ld+json blocks with Organization, ProfessionalService, Person and PostalAddress. The review script strips <script> tags to save tokens, and JSON-LD lives inside a <script> tag, so I handed the panel a version of the site with its schema surgically removed. Gemini reasoned correctly from bad input.

Gemini said there is no llms.txt. False. Both /llms.txt and /llms-full.txt return 200. The models only received rendered pages, so nothing in the prompt could have told them.

Consequence worth noting: section 4, where the panel judged the site's own AI-citation readiness, was partly reasoning about entity signals it could not see. Treat that section as the least reliable of the six. Everything below is verified.

---

Verified against the live site

---

The consensus: the offer architecture is the problem

All five models independently opened with the same thing. Not the design, not the copy quality, not the content. The offer.

The site currently presents two unrelated product lines with eight price points and no explanation of how they relate. A consulting ladder (Free, $500 MiniFix, $4,500 Sprint, $1,500/mo Monitoring) sits directly above a self-serve tool ladder (Free, $49, $129, $299). Nothing on the page says these are different products.

The name collision is the sharpest edge. "Visibility Snapshot" is the free offer on the homepage and services page. On the audit page, "The complete Visibility Snapshot" is the $129 tier, and the same page states "The Visibility Snapshot is not part of the free scan." Both claims are live right now.

The free thing is also called five things: Visibility Snapshot, Free Scan, free AI visibility audit, SEO Audit Report, and "my free audit."

Then the path dead-ends. The nav CTA on every page reads "Get your audit" and points to /seo-audit#pricing, where the highest-intent click in the funnel lands on "Self-serve checkout opens shortly." The contact form's service dropdown tops out at "SEO Audit Report ($0 to $299)." A buyer who wants the $4,500 Sprint has no way to say so.

GPT-5.5 put the cost plainly: it trains the buyer to think this is a small audit-tool business rather than a consultancy. Grok put it as lost intent: hot buyers hit a dead form and downgrade themselves to "just have a question."

---

Kill shot two: the proof is the wrong animal

Three models flagged this independently.

Every hard number on the site is Bone Voyage: DR 0.9 to 62, 4,000 dogs, 31,310 organic keywords, zero ad spend. Those are strong numbers attached to a national dog rescue, and the buyer is a Boulder plumber or dentist.

GLM-5.2 was sharpest on why it does not transfer: Bone Voyage was national, content-driven, and built on original research and press. A local plumber needs map-pack ranking, Google Business Profile work, and city pages. The site never bridges the two, so the buyer concludes "she ranks nonprofits and content sites, not businesses like mine."

Grok noted the homepage also says "Others are clients I'm working with now," plural, when there is one client with no results yet.

The fix nobody has to fake: an explicit page mapping which mechanisms transfer from the rescue to a local service business, and which do not.

---

Kill shot three: the one number she stakes credibility on has three values

Opus caught this and it verifies. Her positioning is "every number in it is measured, never estimated" and "run it like a scientist." The Bone Voyage starting domain rating appears as:

• /case-studies/: "a domain rating of 0.8"
• homepage and /about/: "DR 0.9"
• homepage and /case-studies/bone-voyage-arc/: "DR 12" and "domain rating of 12"

The timeframe is equally unstable: "one focused year," "in about a year," "over five years," "over six years," and a stat block carrying both "Time horizon 5 years" and "~1 year focused SEO sprint" on the same page.

An SEO-literate buyer catches this in about ninety seconds, and the measurement positioning dies with it. This is the cheapest severe fix on the list: pick one baseline, one endpoint, one timeframe, write one sentence, use it verbatim everywhere.

---

What to add, ranked by consensus and return

1. Publish a sample Visibility Snapshot. All five models, unanimous.

The single most-agreed item in the entire panel. She sells a report and nobody can see one. Every objection ("what do I actually get," "is she any good," "is $129 worth it") is answered by the artifact rather than by more copy.

Opus had the best version: run her own top-tier audit against seobeliever.com and publish the whole thing, including the parts where her own site scores badly. A consultant publishing her own bad result is more persuasive than any testimonial she does not have.

Currently /sample-snapshot/ is a 404. Effort: one afternoon, she already owns the tool.

2. Freeze the offer names and fix the intake path. Four models.

One name for the free human review, one name for the machine product, applied everywhere including all 25 blog CTAs. Nav CTA follows the money to the Snapshot, not the audit pricing anchor. Contact form dropdown gains Snapshot, Sprint and Monitoring options and prefills from ?service=. Until Stripe is live, replace "Self-serve checkout opens shortly" with an honest delivery SLA.

3. A public AI-citation scoreboard for her own properties. Grok, and it is nearly free.

Publish the prompts, the engines tested, and whether she gets cited, updated monthly. This is proof of method that needs no client.

Worth adding from my side: she has already built this. The tracker runs weekly and writes to tracking/ai-citations/. Three runs exist. The current honest number is zero organic mentions, and publishing that alongside the method is exactly the "publish your own bad result" move that makes item 1 work. This is the lowest-effort high-credibility asset available, because the hard part is already done.

4. Industry pages. Three models.

Her entire business model is category exclusivity and there is no page mentioning a category. /industries/ is a 404. A Boulder dentist has nothing on this site addressed to dentists.

Opus's version is the strongest: each page carries facts she can gather in an hour with her own tool, such as what the Boulder map pack for "dentist near me" actually looks like today, review counts of the top three, and which assistants name anyone at all. Original, local, current, checkable, and precisely the kind of thing an AI will cite.

5. A market availability table. Opus only, and it is the sharpest unique idea in the panel.

Exclusivity is her only real scarcity and it is currently one sentence buried in a block. Turn it into a table: categories down the side, the six Front Range cities across the top, every cell reading Open except the one she has taken, with a date stamp.

It is honest, verifiable, needs no proof she lacks, and converts "I'll think about it" into "how long is that open." Opus even framed the weakness as the pitch: 42 of 43 slots open is what a three-month-old consultancy looks like, and it means the buyer gets to pick their category first.

6. A week-by-week Sprint breakdown. Two models.

The Sprint is the main revenue product and the vaguest thing on the site. $4,500 for "thirty days of focused execution" with no statement of what happens in those thirty days. A week-by-week scope makes it feel like a build rather than a belief. Two hours.

7. A "new consultancy, old operator" block. Two models.

Neutralise the age objection before the buyer finds it. Both GPT-5.5 and Grok wrote nearly the same copy independently: say plainly that SEO Believer opened in May 2026 and there are no years of client wins, then put the operator proof next to it (adoption.com 1995, Cap Gemini pre-Google internet strategy, Bone Voyage). Grok also noted Cap Gemini is missing from the About timeline entirely, which is a free credibility win.

8. The Houston phone number. Grok only, verified, and awkward.

346-490-0764 is a Houston area code, published as the contact number for a Boulder local-SEO consultancy whose entire pitch is local visibility. A Front Range buyer who notices reads it as "not actually local." Either get a 303 or 720 number, or explain it in one honest line.

---

The one change, if only one

Three of five models converged on the same answer, and it matches the consensus kill shot: make the free Visibility Snapshot the single front door, with one name, one destination, and a contact form that can accept a $4,500 order.

The two dissents are worth knowing. GLM-5.2 argued for rewriting the hero instead, on the grounds that the current subhead talks about her grievances with the SEO industry rather than the buyer's problem. Gemini argued for the name collision alone. Both are inside the same problem, so fixing the offer architecture absorbs them.

---

What is genuinely working

The panel was told not to pad this, and it stayed short.

• Technical execution is strong and several models said so unprompted.
• The writing voice is credible, specific, and free of agency register.
• The 25-post blog library is real depth for a three-month-old domain.
• The market-exclusivity constraint is a genuinely differentiated position, currently underused.
• The refusal to claim client results she does not have reads as integrity rather than absence, and two models called it an asset to lean on rather than apologize for.