Same Homepage, Three AI Engines, Three Different Answers: What We Learned About GEO by Putting Ourselves on Trial
By Alexander Fakeri • June 12, 2026 • Digital Marketing
Category
By Alexander Fakeri • June 12, 2026 • Digital Marketing
By Alexander Fakeri • June 12, 2026 • News, Digital Marketing
There's a comforting myth in marketing right now. It says generative engine optimization (GEO) is just SEO with a fresh coat of paint, and that if you "optimize for AI," you optimize for all of it at once. One playbook, every engine, done.
In June 2026, we tested that myth on the most uncomfortable subject we could find: ourselves.
We asked three of the most widely used AI assistants, ChatGPT, Claude, and Gemini, the same two questions about MOJO Creative Digital's homepage:
The scores came back close: 7/10, 7.5/10 (later revised to 7/10), and 7.2/10. Stop there and you'd conclude the engines agree and GEO is solved. But the scores were the least interesting part. How each engine got to its number was completely different: what it retrieved, what it admitted it couldn't see, and how it behaved when challenged.
| Engine | First score | After pushback | How it read the page | What it admitted it couldn't see |
|---|---|---|---|---|
| Claude | 7/10 | 7/10, with design comments withdrawn | Fetched the live page as text | Visual design, layout, mobile, load speed |
| ChatGPT | 7.5/10 (Visual Design 8.5, Mobile 8) | 7/10 for messaging and positioning only | Answered from indexed text, then checked the live page | Visual design, mobile UX, page speed, CTA placement |
| Gemini | 7.2/10 | 7.2/10 | Automated crawl of text and structure | Said up front it skipped visual rendering |
Anyone can ask an AI to rate their website, and you'll get a confident, tidy answer every time. That confidence is the problem. The first reply is written to sound authoritative whether or not the AI actually looked at anything.
The follow-up question forces each engine to show its work. Did it fetch the live page, or answer from something it had already indexed? What did it actually see, and what did it only infer and then describe as if it had seen it?
Claude wouldn't guess. Asked to rate the homepage cold, it said it couldn't see the page and asked for the URL. Once it had the link, it fetched the page and gave a 7/10: "a competent, well-structured agency site that does the fundamentals right, but it's playing it safe."
It called the client roster the site's single strongest asset and praised the information architecture. It also flagged an oversized news feed on the homepage, some awkward industry copy, and an "Award-Winning" claim in the title tag with no proof on the page.
Under pushback, Claude sorted everything it had said into two buckets: what it actually saw in the retrieved text, and what it had inferred. On the visual design, it was blunt:
"A text extraction tells me nothing about the rendered visual experience... I was pattern-matching from structure and copy, not seeing pixels. That was me overreaching."
Its text-based critique was well grounded. But it rated a visual artifact without seeing the visuals, and only said so when pressed.
ChatGPT gave the most structured first answer: 7.5/10 overall, with a seven-row scorecard. Visual Design got an 8.5. Mobile Experience got an 8, with a quiet note: "(based on structure)."
The critique was useful. It praised the "customer journey architects" positioning and the service breadth, but called the hero "too abstract." It said the page "talks about MOJO before it talks about the client" and has "not enough proof near the top."
Then the pushback landed:
"My prior answer was based on indexed homepage text, not a full visual/browser inspection of the live page. I overstated the certainty of the visual/design assessment."
It had scored the visual design and mobile experience of a page it had never rendered. After checking the live page, it kept its messaging critique, withdrew the design scores, and revised to "7/10 for messaging and positioning."
Gemini's first answer read like an agency teardown: an established agency founded in 2004, strong credibility signals, deep industry segmentation, but a buzzword-heavy hero and a "we serve everyone" problem. It scored the page 7.2/10.
Asked how it formed that opinion, Gemini described a different method entirely. It said it ran an automated crawl that bypassed visual rendering "to focus strictly on the messaging hierarchy, crawlable content, and UX positioning, which is exactly how an optimization tool or a search engine evaluates a site."
Then it went further than the others into hard facts. Claude had noted that the footer lists MOJO's federal procurement identifiers, but Gemini quoted the actual UEI, CAGE, and DUNS numbers. It also pulled the founding year and founder, the SDVOSB certification, the six industry verticals, and the podcast names. Claude described what it read and ChatGPT described how the page felt. Gemini behaved like a system resolving MOJO as a business entity, the way Google's index understands a company.
The three answers line up with how each system reaches the web.
Three indexes, three crawlers, three evidence bases for the same company. Optimizing perfectly for one guarantees nothing about the other two.
There's a flip side, though. Underneath the different plumbing, all three agreed on the substance. Each one independently flagged the same problems: an abstract, buzzword-heavy hero; AI positioning that's claimed but not proven; and a Professional Services line that reads as combative. When three differently built systems land on the same criticism, that's signal, not noise.
That's the real shape of GEO. There's a large shared foundation of clarity, structure, and credibility that every engine rewards. On top of it sits a thinner, decisive layer of platform-specific retrieval that decides whether you show up in a given engine's answer.
We asked three AI systems to rate a homepage, which is a visual thing. All three rated it. And all three, when pushed, admitted they hadn't seen the design.
That reframes the job. When you optimize for AI visibility, you aren't optimizing pixels. You're optimizing the layer these engines actually read: copy, headings, semantic markup, alt text, metadata, structured data, and entity signals. A beautiful design that says nothing extractable is invisible to them. A plain page with sharp, specific, verifiable copy is legible to all three.
GEO is the practice of writing and structuring your content so AI engines like ChatGPT, Claude, and Gemini cite it in their answers. Traditional SEO aims to rank your page in a list of links. GEO aims to make your business part of the answer, and it builds on a strong SEO foundation rather than replacing it.
No. ChatGPT leans on Bing's index, Claude on Brave Search, and Gemini on Google's index and Knowledge Graph. A site that's strong on Google but weak on Bing can do well in Gemini and poorly in ChatGPT.
Generally, no. All three engines rated our homepage, and all three admitted under questioning that they hadn't seen the design. They read text and structure: copy, headings, markup, alt text, metadata, and entity signals.
After it rates you, ask it to show its work: "Walk me through exactly how you formed that opinion. Did you actually access and read the page, what did you see, and what would change your assessment?"
We put our own homepage on trial across ChatGPT, Claude, and Gemini, and we can do the same for yours. Request a quote
MOJO Creative Digital ran this test on its own homepage in June 2026. Quotes from the engines are excerpted from those sessions, with punctuation adjusted to house style. Claude, one of the three engines evaluated here, also helped write this article.
A note on method: this was one session per engine, run in June 2026. AI answers change as models and search indexes update, and the same questions asked on another day could produce different scores and different retrieval behavior. Treat these results as a snapshot, not a benchmark.
