We tested our AEO method on our own site first
Before we sold AEO to anyone, we ran it on karpi.studio. From January to September 2026, AI citations went from 62 a month to 17,484. Visits from AI assistants went from 1 to 49 a month. Eleven inquiries came in that we could trace to AI search, and one of them became a client worth $84,000 a year.
Don't take our word for it. Book a call and we'll open the raw Bing and Google exports with you. →
12,917 AI Citations
Aug 16
January
We started with the questions buyers kept asking on our sales calls. We rebuilt the case studies so each one answers a real question in its first paragraph, and published our first FAQ pages.
February
We answered 22 buyer questions as their own pages, the answer in the first sentence, and rebuilt the five legacy articles that still had search demand instead of replacing them. Two topic hubs opened with their first guides. The curve stayed flat. Groundwork doesn't show in the month you lay it.
March
We published one list, built for LLM retrieval rather than reading: the answer in the first paragraph, every heading a question, every comparison a table. Engines could cite it without rewriting it, and did. For the next three months it earned six of every ten citations the site received and we used that visibility to trade backlinks with other article builders in the space.
April
We took on a glossary project (something that was not on the internet, yet in high demand) and gave every term its own page with a working code example: 281 pages, so a question about any single term has exactly one place to land. Both core service pages were rebuilt in the same shape. Pages cited per day doubled, from four to eight.
May
We published a plain-text map of the site for AI crawlers, added eight guides to the hubs, and anchored 152 source links across 89 glossary pages so every claim points at its evidence. The best single day passed a hundred citations for the first time.
June
We tried to repeat the spring by volume: four more lists and a comparison, in broader categories. Together the five earn fewer citations in a month than the first list earns in a day. The categories were already full of pages saying the same thing, and an engine has no reason to pick a new one.
July
We wrote three narrow lists, each answering a question nobody had a page for, and wired 30 older articles into each other, so a cited page lends its authority to what it links to. The day the lists went live the citations vanished; the lowest day was four. We changed nothing on the site and waited.
August
They came back at 540 a day, four times the July rate, and one of the narrow lists became the most cited page on the site. The glossary earned nearly a fifth of everything, from pages that each get quoted a few times a day. The best day reached 713. Volume from many small pages, not one hit.
September
We published nothing new. Citations kept climbing anyway, because more of the pages already live were being quoted: 37 different pages on a single day, up from four in January, a best day of 735, and a quarter of all citations now landing on the glossary. That was the point of the method. Not one page that hits, but hundreds that each get quoted a little.
What we measure and why
Stage
Metric
Source
Baseline Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
MoM
vs baseline
Visibility
Bing AI citations
Bing
62
52
149
646
2,199
2,500
3,930
12,917
17,484
35%
282.0x
Demand
Branded search impressions
GSC
339
314
394
568
744
704
566
606
616
2%
1.8x
Visits
Generative AI sessions
WF Analyze
1
6
12
6
19
48
46
44
49
11%
49.0x
Pipeline
Attributed inquiries
Karpi OS
0
0
0
0
1
3
1
1
5
400%
new
Sales-accepted leads
Karpi OS
0
0
0
0
1
2
1
1
2
100%
new
Closed-won deals
Karpi OS
0
0
0
0
0
0
1
0
0
-
-
Revenue from sourced leads
Karpi OS
$0
$0
$0
$0
$0
$0
$84,000
$0
$0
Total:
$84,000
Frequency:
Monthly
Quarterly
What worked: The method
Every page published since month 2 follows all seven.
Headings as questions
Each heading is a question a buyer would type. The section under it is the answer.
Answer in the first paragraph
The question the page exists for gets answered before anything else, in words an engine can lift whole.
Tables for anything enumerable
Comparisons, options, steps, prices. Engines lift tables whole and paraphrase prose.
One page per term
Every concept gets its own URL, so a question about it has exactly one place to land.
A plain-text map
llms.txt: who we are, what we do, every key URL with one line on what it answers.
Links between cited pages
Every new page links up to its hub and across to its neighbours. A cited page lends authority to what it links to.
Narrow over broad
A specific question nobody has written the page for beats a broad topic where fifty pages already say the same thing.
Every month, the same loop
Publishing was the visible part. Underneath it, the same five jobs ran every month, driven by what the exports said.
Read the data
Bing and Search Console exports pulled 16 times since May, page by page: which pages were being cited, which were slipping, which queries were new.
Rewrite what the engines already show
Titles and descriptions on the pages with impressions but no clicks: eight pages in May, eight more in June, and the same title bug fixed five separate times.
Link cited pages to the ones that need it
30 articles wired together in one July pass, 152 source links anchored across the glossary in May, and every new page linked up to its hub.
Refresh what loses ground
The pricing article after its citations fell from 132 to 6 in thirty days. The three summer lists two weeks after launch. Nine hub guides in September. 63 of the 65 live articles have been republished with changes since they first went up.
Keep the schema honest
One entity identity across the site, and FAQ markup that matches the visible questions one to one, checked every time a page changed.
What didn't work
× Five list in one week
Broader categories, crowded fields. Three of the five went the whole of the latest month without a single citation.
× The pricing article
Written for search volume, not for buyers: its job was to put the site on the map, and that's all it ever did. It peaked at 132 citations in a month, fell to 6 the next as newer pages on the same question replaced it, and had to be rewritten from scratch. It brought visits. It was never going to bring a lead.
× Two blackouts
Twice, the day after a publish, citations fell to near zero. The second time from 499 to 82 overnight, then 706 four days later. We don't know why. We know not to touch anything while it happens.
9 months, in numbers
The increase in AI visibility translated into measurable business outcomes, from attributed inquiries and a closed customer to $84K in revenue from AI-sourced leads.
37.5k
Total citations which helps with brand visibility
37/day
Cited pages at peak - which makes the brand authority in the space
11
Attributed Inquiries
$84,000
Revenue from AI-sourced leads
AI Citation Performance by Cluster
Cluster
Pages
Citations
Share of total
Top page's share of its cluster
Avg per page
Cluster 1
37
26,908
79.9%
46.3%
727
Cluster 2
96
5,422
16.1%
29.0%
56.5
Cluster 3
27
992
2.9%
38.8%
36.7
Cluster 4
3
132
0.4%
84.8%
44
Cluster 5
3
125
0.4%
94.4%
41.7
Services
2
76
0.2%
98.7%
38
Homepage
1
26
0.1%
n/a
26
Other
4
10
<0.1%
n/a
n/a
Why we measure AI performance in Bing Webmaster Tools and Google Search Console
Only two platforms report how your content actually performs inside AI answers, using their own data rather than estimates.
Bing Webmaster Tools reports citations in Microsoft Copilot, AI summaries in Bing, and select partner products, together with the grounding queries that triggered them. Microsoft does not disclose which partners are included, but the counts come from Microsoft's own logs.
Google Search Console reports impressions in AI Overviews, AI Mode and Discover. It shows which pages appear and how often, but not clicks, queries or citation detail.
Every third-party AI visibility tool works differently. It sends prompts to the models and records what comes back. LLM answers vary from user to user and between identical prompts, so what those tools report is a sample, not a measurement.
That is why the AI performance charts in this case study come from Bing and Google. They are the only first-party sources available. OpenAI, Anthropic and Perplexity have not released publisher analytics, and until they do, there is no other measured data to show.
Questions
How long until AI engines cite a page?
+
Which pages get cited?
+
Why measure in Bing Webmaster Tools and Google Search Console?
+
How do you know an inquiry came from an AI assistant?
+
What does a citation turn into?
+
For a site that's already being crawled, two to six weeks. The glossary was cited within the month it went up. The July lists took eleven days: nothing at all, then nearly the whole of July's citations in a single week. The August list appeared in Bing's report inside a week and in Perplexity inside sixteen days. The exception is groundwork. Schema, FAQs and rebuilt pages showed nothing for six weeks, and everything after them depended on it.
Pages that answer one specific question in their first paragraph. Here that meant the narrow lists and the glossary: one narrow list became the most cited page on the site, and by September a quarter of all citations landed on glossary terms, each quoted a few times a day.
They are the only two platforms that report AI performance from their own logs. Every other tool sends prompts to the models and records what comes back, so it reports a sample, not a measurement.
Every inquiry is logged in Karpi OS with its source. An inquiry counts as AI-attributed only when the visit came from an AI assistant or the person told us that is where they found us.
Visibility first, then visits, then pipeline. From January to September citations went from 62 a month to 17,484, visits from AI assistants from 1 to 49 a month, and eleven attributed inquiries produced one client worth $84,000 a year.