Ninsei LabsLast week we pulled about 950 lines of copy off the four service pages of our own AI agency marketing site. Roughly 42% of the words on those pages. Most of what came out was generic AI-agency boilerplate that read fine in isolation and was indefensible the moment we asked whether it was actually true.
We were not unusually careless. We were unusually willing to look.
This piece is the inventory of what we removed, why it was there in the first place, and a checklist you can run on any AI vendor's site before you sign with them.
The shape of the problem
Most AI agency websites in 2026 are propped up by fabricated proof. Three structural reasons keep this in place.
Generative site builders (Lovable, Bolt, V0, and the WordPress / Webflow templates that came before them) ship with placeholder copy designed to sound compelling. The placeholders include percentages, star counts, project counts, "trusted by" logos, money-back guarantees, and stock-photo testimonials. The defaults exist because the templates were optimized for "looks complete on demo day," not "is defensible in court."
The agency owner ships the site fast. Auditing every placeholder takes hours. The marginal cost of leaving a fake metric in place feels like zero today and only catastrophic in two specific scenarios: a Google manual-review penalty for fabricated structured data, or a lawsuit citing the published guarantee. Both feel hypothetical the morning the site goes live.
The buyer can't tell what's real. They're evaluating five AI agency sites in a tab group. They have no industry reference for which metrics should look outrageous and which should look modest. The fabricated ones often look more credible than the real ones because the fabricated ones are tuned for plausibility.
The result: a generation of AI agency sites where the proof claims are roughly indistinguishable from sales-deck filler.
What we deleted from ours
The full inventory, in the order we found it during the three-phase sweep.
Fake review schema in JSON-LD. The home page was injecting reviewCount: 47, rating: 5 into Google's structured-data layer. There was no review page anywhere on the site. There were no 47 reviews. This is the single most aggressive form of fabrication because Google's structured-data audit can issue a manual penalty against the domain when caught. We removed the schema in Phase 1.
Three fabricated FAQ entries. The "97% lead capture," "3x ROI," and "24/7 support" entries in the page-level FAQ schema. None had a source. All three read like generative-builder defaults and could not be defended if a prospect asked which clients those numbers came from. Replaced with four honest entries describing actual positioning, pricing model, and intake flow.
Hero-section metrics. "97% lead capture," "3x ROI," "20 hours saved per week." Plus a "FREE Implementation Normally $15,000" line that was both a fabricated discount and a discount we had never offered. All four removed.
Problem-section metrics. "$50K+ from manual processes," "67% of teams report ...," "20 hours per week wasted." The cited percentages had no underlying study, and the cost figure was a round number with no methodology. Removed; the section now names four industry gaps (decks instead of systems, chatbots instead of agents, junior hands on senior problems, hourly billing on outcome work) without invented numbers.
Solution-section metrics quadrant. "Zero Lead Leakage" with a trademark symbol on a phrase that was not actually trademarked, plus a four-cell metric grid (3x / 89% / 97% / 24/7) where every cell was a default. Removed. The section now describes the four real Ninsei differentiators: senior judgment, working systems, outcome-based pricing, self-as-proof.
About-section stats. "Lightning Fast / Laser-Focused / Scale-Ready" plus a stats grid showing "50+ projects," "99% client satisfaction," "24/7 support," "3x ROI." The agency had not done 50 projects. There was no client-satisfaction survey to cite. Replaced with the actual Ninsei narrative: the naming story, the four-service breakdown, the senior-judgment-no-junior-hands operating principle.
About-page fabrications. The About page had its own copy: "12+ years experience," "500+ projects completed," "98% happy clients," "25+ industries served." The agency is a year old. We removed these. The page went from 397 to 250 lines.
The full social-proof section. Three testimonials with Unsplash stock photos as the "client avatars." Quotes attributed to people who did not exist. The orphaned sarah-johnson-avatar.jpg asset deleted alongside the section. The site no longer has a testimonials section because we do not yet have testimonials.
Service-page metrics, four pages worth. The AI Automations page had "87%," "95%," "0.3s response time," "50+ languages," "30+ hours saved," "99.9% uptime," plus a "30-day implementation guarantee" that was not in any SOW. The Consulting page had "500+ projects," "285% ROI," "25+ industries," "98% satisfaction." The Micro SaaS page had "50+," "94%," "4-6 weeks guarantee," "340% ROI." The CRM page had six live Stripe checkout links to a SaaS we had already sunset. All removed. Each page also lost its "Book Consultation" CTA in favor of the actual "Book a discovery call" language used everywhere else.
Product-section fabrications. The StyleSuite section claimed "42% increase," "98% uptime," "500+ active users." The Astri section claimed "10,000+ members," "4.9-star rating." Neither product had been audited against those numbers. Removed. Both also carried a "Flagship Micro SaaS" badge, which logically only one product can hold; we changed both to "Built by Ninsei."
URL inconsistency. The StyleSuite product had three different URLs in the codebase: https://stylesuite.io, https://StyleSuite.io, https://stylesuite.app. Three different forms of the same URL across nav, footer, and case studies meant no one was auditing the site. The inconsistency itself is the proof: whatever was on the site, no one was checking.
Across the four service pages alone, the line count went from 2,259 to 1,307. About 42%. The home page and About page lost roughly another 200 lines combined.
What was left
What stayed: the architecture of the site, the four-service framing, the actual case studies (StyleSuite and Astri as built-by-Ninsei products), the honest pricing language (Audit at $5,000, Inbound Agent productized package, custom builds quote-on-request, retainer on request), and the discovery-call CTA. The site is shorter, slower-rhythmed, and reads like an operator's site rather than a deck.
We added the Neuromancer naming origin to About, four-service descriptions sharpened to match the underlying offerings doc, and one explicit honest stage statement ("we're a year old, here's what we've shipped").
A checklist for evaluating an AI vendor's site
Run this when you are evaluating an AI agency, a micro SaaS shop, or any AI-services vendor.
Look for percentages in the hero section. If you see "97% accuracy," "3x ROI," or "10x faster" without a source, ask which study. Specific numbers without specific provenance are almost always defaults the agency owner forgot to remove.
View-source the page and search for aggregateRating or reviewCount in the JSON-LD. If you find a numeric rating with no corresponding testimonials page, the rating was invented for the schema. Google has manually penalized domains for this. The presence of fabricated schema on a vendor's own site is a signal about how they handle proof claims elsewhere.
Reverse-image-search any testimonial avatar. If the photo is from Unsplash, Pexels, or any stock library, the testimonial is fiction. This is the single fastest test and most vendor sites still fail it.
Read every "X-day guarantee" and "Y-day money-back" line. Ask whether the guarantee is in the SOW or the terms of service. If the guarantee is not contractually written, it does not exist, and the vendor is using it as a closing prop.
Count the case studies. If the site says "500+ projects" or "25+ industries served," ask for the project list. A vendor with 500 real projects can name 20 of them without thinking. A vendor with 5 real projects will be hand-wavy.
Check for trademark theater. The TM symbol applied to phrases like "Zero Lead Leakage" or "AI-Powered Pipeline" usually indicates a builder that learned the symbol adds gravitas. Real trademarks are searchable in the USPTO database.
Click through every link in the navigation and footer. URL inconsistency (product.io versus product.app versus Product.io) means no one is auditing the site. The substantive claims on a site no one audits are unlikely to be more reliable than the URLs.
Finally, look for what is not there. An honest case study includes a "what didn't work" paragraph. An honest pricing page either fully discloses or honestly defers. An honest about page includes the founding date. Most fabricated AI agency sites strip these out because they collapse the proof claims around them.
The take
The exercise is not about virtue. It is about defaults. The defaults of every modern site builder are tuned for "looks complete," not "is defensible." Most AI agency owners ship and do not audit. We did the same thing on our first pass.
The buyer-side asymmetry is the part worth naming. The agency owner has a fast, cheap, one-time decision to either audit or not. The buyer has a slow, expensive, contractual decision based on what the agency owner chose to leave on the site. The cost of fabrication falls on the wrong side of the trade.
We have to assume most AI vendor sites have not had this audit run. Run the checklist on the next two you're evaluating. The results will tell you something about how the vendor treats truth claims in the rest of their work.