Does AI SEO actually work? Receipts from a live site

Does AI SEO work? Yes — when the AI makes bounded, verifiable changes and someone honestly measures the results. No — when it means mass-generating pages and hoping. The interesting question was never really "does it work," it's "how would you know?" Most of this industry, human and AI alike, is structured so that you can't. This post is about the measurement standard we think any AI SEO vendor should be held to — and how we apply it to a live site.

Full disclosure as always: we build Cliff, an AI SEO employee, so we have an obvious interest in the answer being yes. That's exactly why we run it on our own sites first and publish the standard we want to be judged by.

The receipts philosophy

Here's the core idea, and you can steal it to evaluate any vendor:

Every SEO change should be logged with three things: what was changed (with the before-state), what outcome it's expected to produce, and a date when that expectation gets checked. When the date comes, the result gets reported — including when the answer is "nothing happened."

We call these receipts. A receipt is different from a report. A report says "we did 14 optimizations this month and traffic is up 8%" — activity next to an outcome, with the connection implied. A receipt says "on the 12th we rewrote this title tag; we expected CTR on this page to improve within 30 days; here's the before and after" — one specific claim you can check, and one that can visibly fail.

Receipts change the vendor's incentives, which is the actual point. If every change carries a testable prediction, you can't hide low-value busywork inside an activity count. Some receipts will show misses — real SEO has misses — and a vendor whose log shows zero misses is a vendor whose log is marketing.

This standard is also, frankly, easier for an AI to meet than a human team. Logging every change with its before-state, scheduling a day-30 check, comparing search data before and after — that's tedious bookkeeping for a person and trivial for software. It's why we think honest measurement will end up being the agent category's biggest advantage, more than speed. It's built into how Cliff works at every tier.

The live proof: webflow.jobs

Talk is cheap, so here's the site we point at: webflow.jobs — a job board for the Webflow ecosystem that we operate. It's not a client, it's ours, which means we can show it without asking permission, and every SEO decision on it since we turned the loop on has been made by the agent: recurring audits, metadata rewrites, internal linking, content optimization, technical fixes — each change emailed as it happened, each with a logged before-state, each checked against an expectation afterward.

What we can say plainly: since the autonomous loop took over, webflow.jobs has reached record organic traffic, and the site converted its first paying customers while the agent ran its SEO. No human was doing weekly SEO work on that site — the humans reviewed and occasionally vetoed.

What we won't do is dress those sentences up with invented percentages. The specific numbers belong in receipts (above, once published), where you can see the change they're attached to — not free-floating in a blog post.

What "working" actually means

When we say AI SEO works, here's the definition we're using — and the one worth demanding from anyone:

• Impressions: the site appears for more queries, more often. The earliest reliable signal, usually visible in weeks.
• Positions: specific tracked pages move up for specific tracked queries. This connects changes to outcomes.
• Clicks: more actual visits from search. This is the one that pays.
• Not: an internal "SEO score" going from 61 to 88, audit-issue counts falling, or word counts rising. Those measure activity inside the tool, not results in the market. Vanity metrics aren't lies, exactly — they're just answers to a question nobody asked.

Timeframes matter too. Honest AI SEO shows impression movement in weeks and meaningful click movement in months — the same physics as human SEO, because it's the same search engine. The speed advantage of an agent isn't faster Google; it's that the work actually happens every week instead of when the retainer gets around to you.

One more metric worth adding in 2026: AI citations. A growing slice of buyers never see a results page — they ask ChatGPT or Perplexity and read an answer. Whether your site gets named in those answers is now part of what "working" means, and it's trackable the same way rankings are: a fixed set of buyer questions, checked on a schedule, trended over months.

When AI SEO fails

It genuinely does, and the failure modes are predictable:

1. Content spam at scale. The most common one. Generating hundreds of thin pages is the one strategy Google explicitly targets (scaled content abuse), and AI made it cheap enough to be tempting. Any agent whose main output is "more pages, fast" is running this play. Bounded optimization of existing pages is a different risk class entirely.
2. No feedback loop. Tools that change things but never check search data afterward are guessing serially. If the system can't tell you what happened after its last change, it isn't learning.
3. Nothing worth ranking. If a business has no pages that deserve to win — no real product, duplicate-of-everyone content, no differentiation — AI executes the optimization of nothing. An agent multiplies what's there; it can't multiply zero. (A human agency can't either; they'll just take longer to tell you.)
4. Unbounded autonomy. An AI rewriting whatever it wants, including prices, claims, and navigation, will eventually write something wrong at the worst possible place. This is a design failure, not an intelligence failure — scope caps and approval gates exist precisely because the tail risk of "rarely wrong but unbounded" is a live website saying something false. Every change Cliff ships is scoped, recorded, and reversible in one click for exactly this reason.
5. AI answers replacing your clicks. The uncomfortable one: for some informational queries, AI engines now answer directly and the click never happens. That's not AI SEO failing, it's the terrain changing — and it's why optimizing to be cited by AI engines is now part of the job. We wrote the practical playbook in our generative engine optimization guide.

How to evaluate any AI SEO vendor in ten minutes

Ask these, in order, and watch how they're answered:

1. "Show me the change log for a real site." Not screenshots — the running log with before-states.
2. "Show me a change that didn't work." The most revealing question in the industry. Honest vendors have misses on file; dishonest ones have adjectives.
3. "What can it change without approval, and what can't it?" You want crisp boundaries, stated instantly.
4. "How do I undo something?" The answer should be one click, not a support ticket.
5. "What do you measure, and where?" Right answer: impressions, positions, clicks, in your Search Console — data you can verify without them in the room.

Any vendor in this category — us included — should pass all five without flinching. If you're still mapping what these products even are, start with What is an AI SEO agent?, and if you're weighing the cost against human alternatives, How much does SEO cost in 2026? has the sourced numbers.

FAQ

Can AI really do SEO on its own? It can execute the recurring 80% on its own: audits, metadata, internal linking, content optimization, technical fixes, reporting. It shouldn't operate unbounded — new pages, structural changes, and anything customer-facing that states a fact deserve human approval. That division isn't a limitation of current models; it's just good engineering.

How fast does AI SEO show results? Same physics as human SEO: impression movement in weeks, ranking movement in one to three months, meaningful click growth in three to six. The agent's advantage is consistency of effort, not a faster algorithm.

Will Google penalize AI-driven SEO? Google's spam policies target scaled content abuse and manipulation — not the method of production. Optimizing existing pages, fixing metadata, and improving internal links are the same changes a human SEO would make, made more consistently. The risk concentrates almost entirely in mass content generation, which is why bounded agents avoid it.

What's the single best proof a vendor can show? A dated change log on a live site with before-states, predictions, and outcomes — including misses. Everything else is a pitch deck.