AI SEO case study: our own site, 61 new pages, four months of data.

1. The site, the windows, and what counts as a result.

The property is theseoagent.ai, registered in Search Console as a domain property so subdomains and both protocol variants roll into one number. It is a young site: the first marketing pages went up in late April 2026, and through the first half of May the whole domain took fewer than 100 impressions a week. That matters, because a case study on a fifteen-year-old domain measures something different from a case study on a site with no accumulated authority to lean on.

Two windows, 28 days each. The recent one runs 20 August to 17 September 2026, which is the last full window available at the time of writing. The earlier one runs 21 May to 17 June 2026, the same 28 days one quarter back. Twenty-eight days is the floor worth using: Search Console finalises on a two to three day lag, so the freshest days in any export are partial, and anything shorter than a whole number of weeks gets distorted by the weekday traffic curve. Both windows were fixed before the export ran, which is the only defence against picking the window that flatters the chart.

The result metric is clicks. Impressions are reported, because leaving them out would hide the shape of the story, but they are not the win. Average position is reported as context only: Google documents it as an impression-weighted average, so a site-wide figure is a blend of everything the property was eligible for, not a ranking anyone holds. The same separation applies when you build a monthly client report: one column for exposure, one for demand captured, never a blended growth figure.

2. The two windows, side by side.

Here is the whole result in one table. Every figure is a Search Console export for the property above, taken on 19 September 2026.

Month by month, the same property: May took 20,492 impressions and 17 clicks. June, 67,729 and 52. July, 71,702 and 139. August, 158,932 and 281. The first seventeen days of September took 98,031 and 259, which is the fastest stretch in the series. At the weekly grain the low point inside the window was 10,893 impressions in the week beginning 15 June and the high point was 42,968 in the week beginning 7 September.

Read the table in the right order and it says something less flattering than the headline. A 2.5 times lift in impressions on a site that started at an average position of 72.9 is mostly a statement about how many more queries the site became eligible for, not about how many it won. The number that carries weight is 36 clicks to 384, and the number that explains it is 72.9 to 29.8. If you want to model the next four months from this base, the forecasting method takes the position assumption as its main input for exactly this reason.

3. Three levers, all pulled at once.

This is a case study, not a controlled experiment. Three things changed across the window, and low-volume outreach for external links ran alongside all of them. Nobody can cleanly attribute the result to one lever, and any case study that claims otherwise from a period like this is guessing with confidence. What follows is the full list, dated, including the parts that probably did less than they look like they did.

Lever one: publishing cadence, on rails.

On 1 June the site had 34 indexable blog and feature pages. On 19 September it had 95. That is 61 new pages in roughly fifteen weeks, an average of about four a week, each one pulled off a keyword queue by a scheduled job rather than chosen by whoever had an idea that morning. Every page is built to a fixed per-shape recipe and has to pass a validator before it can ship: required sections, a minimum FAQ count, schema present, images resolving, no invented internal links. That is the same SEO automation pipeline we sell, pointed at our own domain, which is the only reason we are allowed to publish these numbers at all.

Lever two: clusters instead of one-off posts.

The 95 pages are not 95 independent bets. They sit in eight clusters, each with a hub page that owns the head term and a set of members that own the long tail around it. A cluster ships as a unit: hub first where one does not exist, then members, then the links between them. That structure is the difference between building topical depth on purpose and producing a blog archive that happens to be large.

Lever three: an internal-link check that blocks the push.

This one shipped on 24 August and is the change we would keep if we could only keep one. Every new page has to be registered in a cluster map, carry at least three contextual in-body links out to distinct pages, and receive at least three contextual in-body links from existing pages. A graph checker runs before the commit lands and exits non-zero on an orphan, an under-linked page, or a dead end. Footer and nav links do not count toward the threshold, because boilerplate links carry almost no weight and counting them is how a site convinces itself it is internally linked when it is not. At the time of writing the graph holds roughly 1,900 contextual in-content links across 132 indexed pages, with zero broken links, zero dead ends and zero under-linked pages. The handful of orphans left are named in the report rather than quietly excluded from it. If you want the same discipline without the build step, our internal linking tool does the audit half of it.

4. Impressions moved months before clicks did.

This is the finding worth stealing, and it is the one most case studies quietly hide. Between May and June impressions went from 20,492 to 67,729, a 3.3 times jump inside a single month. Clicks went from 17 to 52. In absolute terms that is 35 extra clicks against 47,000 extra impressions. Anyone reporting a blended growth number in early July would have declared a win. The site was not winning. It was becoming eligible.

The explanation is in the position column. Average position across the property was 72.9 in the earlier window, which is page eight. A page on page eight is shown for long-tail variants, registers the impression, and is never seen by a human who scrolls. The click curve is brutally front-loaded: Backlinko measured position one at 27.6% of clicks across roughly four million results, with the share collapsing across the rest of page one and effectively vanishing after it. Everything below the fold of page two is exposure, not traffic.

Clicks tracked position, not volume. They started moving in mid-July, when the first pages published in May and June began crossing into the top 30: 14 clicks in the week beginning 15 June, 42 by 20 July, 65 by 17 August, 125 by 7 September. Across those same weeks impressions grew about four times over. Clicks grew about nine. The lag between an impression curve and a click curve is the single most useful thing a four-month window teaches you, and it is why a three-week pilot is worthless for judging whether a steady publishing cadence is working.

5. The cluster that carried most of it.

Of the eight clusters, one produced a clearly outsized share of the movement: the set of pages about how AI search engines pick their sources. The first page in it, a ranked roundup of GEO tools, went live on 8 July. In the recent window it alone took 22,247 impressions and 17 clicks at an average position of 26.9, and it now receives 25 contextual in-body links from elsewhere on the site. The cluster hub, our GEO agent page, did not exist until 21 August, six weeks after the page it was supposed to support. Building the member before the hub is the wrong order, and the cluster took longer than it needed to as a result.

The cluster worked because the demand is genuinely new and the competition for it is thin, not because the pages are better written than the rest. Anyone entering the same territory should read how generated answers pick their sources before writing anything, and treat direct-answer formatting as the on-page requirement it has become. Timing was a bigger factor than craft here, which is an uncomfortable thing to publish and also the truth.

The other clear winner was the homepage, which is the page the whole internal-link structure points at. In the earlier window it took 1,031 impressions and 19 clicks at an average position of 38.0. In the recent one, 7,078 impressions and 132 clicks at an average position of 8.8. It now carries 78 contextual in-body inbound links, and across the recent window the property averaged position 6.3 for the head term this whole product is built around. Nothing about the AI SEO agent homepage changed materially in the window. The pages pointing at it did.

6. The numbers we refuse to count.

The 165,283 impressions figure overstates the result, and it is worth showing exactly where. Three pages account for a large share of the total while contributing almost nothing a business could use.

1. A product review page: 27,691 impressions, 3 clicks, average position 6.7. Position 6.7 is genuinely page one. The query behind it is someone typing a product name to reach that product. They are not looking for a review and they do not want us. Ranking well for a navigational query you cannot satisfy produces a beautiful chart and no customers.
2. A tool roundup: 16,627 impressions, 2 clicks, average position 67.1. Page seven, high-volume head term, nothing. This page is either going to climb or it is going to sit there forever collecting impressions that flatter a monthly report. We do not know which yet, and pretending otherwise would be the dishonest part.
3. A generic alternatives page: 13,878 impressions and zero clicks in the earlier window, at average position 76.0. That page contributed a fifth of the impressions in the before window and nothing else. It is why the earlier window shows 66,632 impressions against 36 clicks, and why the before number is less impressive than it looks.

There is a second reason the impression column is getting less trustworthy for everyone, not just us. Pew Research Center instrumented US search sessions and found a traditional result was clicked on 8% of visits where an AI summary appeared, against 15% where none did. An impression on a query that triggers a generated answer is worth materially less than the same impression was two years ago, and no dashboard adjusts for it. If you want the counted-citation view of that shift, the AI search statistics reference keeps the sourced numbers in one place.

7. A worked example: one page, before and after.

Take a single page through both windows. Our ranked roundup of automation tools went live on 22 May 2026, three weeks before the start of the earlier window. It is a useful specimen because it existed for both measurements and its ranked list never changed.

Position 78.4 to 21.1 on a page whose body copy barely moved. The two metadata passes did some of that work. The rest came from the site around it: the page now receives 55 contextual in-body inbound links, and the large majority of them come from pages that did not exist on 17 June. That is the compounding effect the internal-link rule is built to force, and it is the clearest signal in the entire dataset that publishing into a structure beats publishing into an archive.

Now the honest reading. Nine clicks in 28 days is not a business. Position 21.1 is still page three. Four months in, this page is on a trajectory, not at a destination, and the base rate says that is normal: Ahrefs found only 5.7% of newly published pages reach a Google top 10 for any keyword within a year. A case study that stops at the percentage lift and never prints the absolute click count is hiding this paragraph.

8. How to run this study on your own site.

The procedure takes about two hours if your Search Console property is already verified, and it produces something a sceptical reader can check. Six steps.

1. Fix both windows before you look. Equal length, one recent, one an equal distance back. Write the dates down first. Choosing the window after seeing the chart is the most common way a flat quarter turns into a case study.
2. Export both at page and query level. Site totals hide everything worth knowing. The page export tells you what moved; the query export tells you whether the movement came from queries a buyer would type or from navigational traffic you cannot serve.
3. List every change, with dates. Pages published, metadata rewritten, links added, technical fixes, placements earned. Include the ones you suspect did nothing. A site audit run at the start of the window gives you a clean baseline to date the list against.
4. Split clicks from impressions. Two columns, never blended. If the impression line moved and the click line did not, say so in the same paragraph rather than three sections later.
5. Name the losers. Pull the pages with the highest impressions and the lowest click-through and print them. This costs nothing and buys more credibility than any of the wins.
6. Publish the table. Impressions, clicks, click-through rate, average position, both windows. A reader with their own export should be able to rebuild your table from scratch.

The part most teams cannot sustain is not the measurement, it is the supply. Sixty-one pages in fifteen weeks is about four a week of finished, researched, cited work, which is a full-time writer and an editor at agency rates. Ours came off a pipeline that validates keywords against live search data, writes the draft, checks the claims against real sources, wires the internal links, refuses anything that fails the quality gate, and publishes straight into the CMS. That runs at $99 a month flat, which is the number that makes 61 pages a decision rather than a budget round.

9. Common mistakes in SEO case studies.

Six patterns show up in almost every case study that turns out to be unverifiable. The first three are errors of method, the last three are errors of omission, and the omissions do more damage.

1. Reporting impressions as the result.

Impressions are the easiest number in the dataset to move and the least connected to revenue. Our own June proves it: a 3.3 times impression jump alongside 35 extra clicks. If a case study leads with impressions or with a blended growth percentage, assume the click line was flat and look for it.

2. Choosing the window after seeing the data.

Any four-month series contains a flattering slice. Fixing both windows in advance costs nothing and is the only structural defence, which is why it belongs in the write-up as an explicit sentence rather than an assumption.

3. Claiming one cause for a multi-lever change.

We changed cadence, structure and internal linking at once, and did some outreach throughout. The correct sentence is that all of it happened and the result followed. The tempting sentence attributes it to whichever lever the author is selling. Both sentences are the same length.

4. Quoting a site-wide average position as a ranking.

Average position is weighted by impressions, so publishing a batch of long-tail pages can drag it down while every individual page improves, and retiring a low-position page can lift it while nothing gets better. Use it for direction across matched windows. Never present it as a rank.

5. Omitting the pages that lost.

Sixty-one new pages always produce some that draw nothing, and a study with no losers in it has been curated. Listing them is cheap, and it is the fastest way to signal that the rest of the numbers were not curated either.

6. No dates, no property, no export.

A cropped screenshot of a rising line is unfalsifiable by construction. Name the property or at least the vertical, give both date ranges, and print the table. If the client will not let you name them, the numbers and the dates still work; the anonymity is not what makes a case study weak.