Three Months After Rebuilding Our Site for AI Search: The Real Numbers

This is Part 4 of our AI search experiment. Part 1 set the baseline and the checklist, Part 2 measured the first 31 days, and Part 3 documented the layer that serves every page as Markdown to AI agents. This part reports three months of Google Search Console data and four weeks of Google Analytics 4 data, checked on 20 September 2026 — including the numbers that did not go our way.

The three-month numbers

Two cautions on reading this table. Average position blends every query, date, country and device into one number, so 28.7 is not where any single page ranks. And only one of the 30 clicks is visible in Search Console's query table; the rest sit behind queries Google does not disclose, which is normal for a small site.

For scale: in Part 2, the first 31 days after the rebuild (8 June–8 July) produced 479 impressions and 11 clicks. The new window overlaps that period by about three weeks, so it is not a clean before-and-after, but it averages roughly 2,100 impressions a month.

Which pages Google showed

The New Zealand cluster carried the experiment: the two NZ agency guides, the NZ cost guide and the Webflow NZ service page account for 3,624 of the 6,316 impressions, about 57%. The pages we built for specific buyer questions in a specific market did far more than the generic ones.

What showed up in Google's AI features

Search Console now reports impressions inside Google's generative AI features separately. For the same three months we had 259: Superside alternatives 84, the Webflow agencies NZ guide 83, the Norway agencies guide 20, the homepage 13 and the Webflow NZ service page 12. These are impressions inside Google's AI answers — not clicks, not ChatGPT traffic and not leads — and we do not add them to the Web total, because the same searcher can be counted in both.

AI assistants now send more visits than organic search

In the 28 days from 23 August to 19 September, AI assistants sent 49 sessions against 26 from organic search, and 31 of those 49 were engaged sessions (63%, against 19% for Direct). Where they landed is the interesting part: the ELCora migration case study took 10 sessions and VetaDerm 8 — 18 of 49, about 37% — followed by the industries hub with 5 and the Dubai real estate, Glasshouse and Profrakt case studies with 3 each. Assistants send people to proof, not to blog posts about proof.

One honest caveat: some of that traffic is us. GA4 shows chatgpt.com and a claude-test source on the homepage in the last week of the window, and our own checks land there too. We have not yet separated our visits from real prospects, so not every AI session is a potential client.

What worked

• Market-specific buyer guides: the New Zealand pages produced 57% of all impressions, and "which webflow agencies operate in new zealand" averaged position 7.2.
• Comparison pages: Superside alternatives earned 1,102 Web impressions and the most AI-feature impressions of any page (84); "what should i use instead of superside" averaged position 7.9.
• Startup-specific comparisons: "designjoy alternatives for startups" averaged position 8.7 and "designjoy alternative" 13.6, with the site's only visible query click.
• Engineering write-ups: the Python to Rust migration post has the best average position of any page (11.7) and drew 45 GA4 sessions across all channels in four weeks.
• Case studies as AI landing pages: ELCora and VetaDerm are where AI assistants send people.

What did not work (yet)

• Clicks: 6,316 impressions turned into 30 clicks. Most impressions come from positions 20–50, where almost nobody clicks — the NZ agencies guide had 808 impressions and no clicks at position 50.7; the NZ cost guide 742 and none at 41.5.
• Head terms: "webflow agency new zealand" (206 impressions, position 19.0) and "superside alternatives" (185, position 26.2) are stuck on pages two and three.
• Leads: GA4 recorded one generate_lead event in the four weeks, from a Direct session that landed on the contact page. None came from AI assistants or organic search, and we have not yet matched that event to a real inquiry.
• Indexing: 28 URLs sit in "Discovered – currently not indexed", including four commercial Webflow pages — Webflow development, WordPress to Webflow, Wix to Webflow and Webflow to Next.js — even though all four return 200, are self-canonical and are set to index, follow.
• Speed evidence: Search Console has no Core Web Vitals field data for the site yet, so we cannot claim real-user speed is good.

What we are doing next

• Put a concrete offer on the pages people already visit — the Python to Rust post, the ELCora and VetaDerm case studies — and measure CTA, form start and confirmed inquiry separately.
• Turn the DesignJoy alternatives guide into a comparison of real providers with verified prices and check dates, instead of creating another page for the same query.
• Ask partners and clients for independent confirmation of our New Zealand work, and for honest Clutch reviews without any reward.
• Inspect the four unindexed commercial URLs in Search Console, add unique proof where it is missing, then request indexing and re-check in two to four weeks.
• Count leads properly: track CTA, Telegram, email and calendar clicks with their source page, and separate our own test visits from real prospects.

What we are deliberately not doing: mass-producing industry or city pages, buying links or a GEO subscription, or rewriting markup for AI. Google's own guidance on AI features says sites do not need new machine-readable files, AI text files, markup or Markdown to appear in its generative AI results, and that no special schema is required. We keep llms.txt and Markdown because other agents use them, not because Google rewards them.

What three months do not prove

Three months shows direction, not causation. Several changes shipped together, the GA4 and Search Console windows differ, and the AI Assistant channel still includes our own checks. The single key event is not a verified lead. What the data does show is that clear, market-specific buyer pages get seen by Google and its AI features, that AI assistants send engaged visitors to case studies, and that the next bottleneck is turning visibility into clicks and inquiries — a conversion problem more than a content-volume problem.

If your company has strong work but search and AI systems struggle to describe it, this is the kind of system we build through AI SEO & AEO: entity clarity, answer-ready buyer pages, internal links and a measurable path from discovery to contact. For the baseline and checklist, start with Part 1; for the first 31 days, read Part 2.