The AI Citation Index
This is the instrument the Sprint runs, pointed at our own category and left running in public. Every week it asks ChatGPT the questions a real buyer would ask and records exactly which vendors get named, which sources the answers cite, and how much the list moves between runs. Three dated entries so far. The same instrument, pointed at your category, is what you buy.
- Who ChatGPT actually recommends in this category, question by question, dated: iPullRank, Siege Media, First Page Sage, Profound, Peec, Otterly and others, with the exact query that produced each name.
- How much the answers move. Two samplings of the identical query on the same day agreed on only two of six names. One check gives you one draw from a list that changes between runs.
- Which sources the answers draw on. arXiv appeared in 11 of 22 ChatGPT query samplings across entries 2 and 3. The mix moves: the first-party share of cited sources ran 43 percent on 2026-07-16 and 69 percent on 2026-07-23, so which class of source carries an answer is itself something you measure rather than assume.
- Our own row, at zero, published first. Deep Ocean has been named on 0 of the 11 tracked queries in every entry. We print our own number before anyone else's, which is how you know the rest of the table is not decorated.
Named for the category query
The tracked query for this category: “Best AI search optimization agencies 2026”. These are the domains ChatGPT named in its answer.
| Domain | Named in |
|---|---|
| iPullRank | Comparison table + shortlist |
| First Page Sage | Comparison table + shortlist |
| Siege Media | Comparison table + shortlist |
| Omniscient Digital | Comparison table |
| Optimist | Shortlist |
| NoGood | Shortlist |
On one of the buying questions this panel tracks, the answer named six vendors from our category and did not name Deep Ocean. Our own row follows.
The full week-1 panel, all 8 questions
The category query above is one of 8 tracked buyer questions sampled this entry. This is the complete panel, the same data published in our own white-label evidence-report sample, question by question.
| Tracked buyer question | Named? | Named instead |
|---|---|---|
| Best answer engine optimization (AEO) agency to get a brand cited by ChatGPT and Perplexity | No | LoudFace, XQL Group, First Page Sage, Siege Media; platforms: Profound, HubSpot AEO, Semrush AI Visibility Toolkit, Peec AI, Otterly.ai |
| Who can fix my brand being invisible in AI answers even though we rank well on Google? | No | Profound, Goodie AI, Omniscient Digital |
| AI visibility audit that shows which competitors ChatGPT recommends instead of my brand | No | Profound, Goodie AI, Peec AI, Scrunch AI, Otterly.AI |
| Agencies that offer a fixed price AEO install with a measurable citation guarantee | No | Revlift, AEO Engine, UnFoldMart, AEO Agency USA. The answer itself called this offer shape "still a gap in the market" |
| How do I make my website readable and citable by AI assistants, and who offers this as a done-for-you service? | No | No vendor named in the captured answer body |
| Best AI search optimization agencies 2026 | No | iPullRank, First Page Sage, Siege Media, Omniscient Digital, Optimist, NoGood |
| White label AEO or GEO delivery partner for marketing agencies | No | Blobic, AEO Engine Partner Program, Nico Digital, AI GEO Insight Agency |
| How can my SEO agency add answer engine optimization services without hiring in house? | No | No vendor named in the captured answer body |
Sampled 2026-07-08, week 1, panel v2, single engine (ChatGPT web session). Same data, white-label format: see the evidence-report sample.
Recurring across the wider panel
The category query above is one of 8 fixed queries sampled this entry. These domains were named or cited across more than one of the 8, not just the category query:
| Domain | Type |
|---|---|
| LoudFace | Agency |
| XQL Group | Agency |
| Profound | Platform |
| Otterly | Platform |
| Peec | Platform |
On the fixed-price question, the engine stated it could not verify any provider combining a one-time fixed price with a contractually measurable citation guarantee, and in an earlier entry called that offer shape “still a gap in the market”.
Captured on the tracked query, entries 1 and 3. We publish it because it cuts both ways. Deep Ocean sells the fixed price and deliberately does not sell the citation guarantee: we retired that promise because we have not found a published, controlled demonstration that an intervention changed which brand an AI answer names, ours included. The engine has identified a real gap, and we are filling the half of it we can stand behind. A vendor selling you the other half today is selling a result nobody has shown.
Our own row, entry 1
deepoceanstudio.com was named on 0 of the 8 tracked queries in this sample. We publish our own number before anyone else's. This is the panel v2 baseline; the current row is entry 2 below, 0 of 11 on panel v3, which is the number quoted on our homepage.
Entry 2: the first panel v3 row
Panel v3 keeps the 8 vendor-recommendation questions from entry 1 and adds 3 data-source questions the owned research lane can plausibly earn a citation on. We sampled the full 11-question panel twice on the same day, in two separate sessions, to see how stable the answers are.
Our own row, entry 2
deepoceanstudio.com was named on 0 of the 11 tracked queries. Both independent same-day samplings agree, and a parallel Bing retrieval check found the domain absent on all 11 as well. We publish our own absence first, now on the wider panel.
What we did between entry 1 and entry 2
One action on our own domain sits between the two rows, and the result was a null. We submitted a GoodFirms directory profile on 2026-07-08 and it went live on 2026-07-09. In entry 2, GoodFirms was not among the cited sources in either same-day sampling, and no directory profile appeared in any cited-source set on any of the 11 queries. That action predates our preregistered flip protocol and was never registered under it, so it could not have counted as a flip in either direction. We record it here because this index annotates every action we take on our own domain, including the ones that produce nothing.
What repeated sampling showed
Across the two same-day samplings of the identical category query (“Best AI search optimization agencies 2026”), only two names appeared in both answers: iPullRank and Siege Media. Every other agency named in one answer was absent or replaced in the other. This is why the index samples each query more than once and will not call any single answer “who wins”: the vendor lists move run to run, so a one-shot list would overstate what is actually stable.
The three data questions
On the added data-source questions, the engine's answer now matches judgments Deep Ocean has already published: that publishing an llms.txt file shows no measurable citation lift, and that most AI citations, on the order of 84 percent or more, come from earned third-party sources rather than a brand's own site. The engine credited those points to other sources, among them arXiv, Muck Rack, and Chrome for Developers, not to us. Getting cited as the source on these questions is what the owned research lane works toward, and this row is its baseline.
Entry 3: the third row, sampled on time
Same 11 questions as entry 2, plus the 3 control questions we sample alongside them. The controls are category questions where we have run nothing and expect nothing, so that ordinary week-to-week movement in the answers is not mistaken for something we caused.
Our own row, entry 3
deepoceanstudio.com was named on 0 of the 11 tracked queries. The controls behaved as controls: stable, category-relevant, no sign of an engine-wide shift. Third row, same number.
Limits of this row, stated before you ask
Entry 2 was sampled twice on the same day; entry 3 was sampled once per question, so it carries less protection against a single volatile answer. The Bing retrieval check we normally run alongside the panel was not run this week. The reasoning-effort setting was High, where entry 1 ran on the faster default, and we now record that setting on every row rather than assume it. None of that changes the number, and we would rather print the weaknesses than let a third row look sturdier than the two before it.
Who was named instead, entry 3
On the category question the answer named iPullRank, Siege Media, Optimist, Directive, Go Fish Digital and Omnius, and cited Clutch among its sources. Across the rest of the panel the recurring names were iPullRank, Amsive, NoGood, Omnius, Brainlabs, Single Grain, and on the agency question Butter Marketing, Agency Platform, Big Fin SEO, Nico Digital and Blobic. Two questions produced no vendor at all: on one, the engine offered to run the visibility audit itself instead of recommending anyone; on the other, it answered with method only. On the fixed-price question it stated it could not verify any provider combining a one-time fixed price with a contractually measurable citation guarantee.
The three data questions returned the same pattern as entry 2. The engine repeated judgments we publish, that roughly 85 to 90 percent of AI citations come from third-party sources rather than a brand's own site, that there is no solid evidence an llms.txt file increases citations, and that citation share has to be treated as a repeated-sampling problem rather than a one-time check. It credited Muck Rack, Foundation, arXiv, llmstxt.org, OpenAI and Google. It did not credit us.
Which sources the answers draw on, aggregated
Each entry above records the cited sources answer by answer. Aggregating them across questions and engines shows which sources recur. The weekly meter stays ChatGPT only; the Perplexity and Bing figures come from a dated capture run of the same 11 questions, kept to the same receipts discipline.
| Recurring source | Where it recurs | Class |
|---|---|---|
| arXiv | 11 of 22 ChatGPT query samplings, entries 2 and 3 | Academic preprints |
| OpenAI, Google and Microsoft documentation | 6 of 11 ChatGPT queries, entry 3 | Engine and search docs |
| iPullRank | 4 of 11 ChatGPT queries, entry 3 | Agency site, also named in answers |
| AEO Engine | 2 ChatGPT queries, entry 2, and 2 Perplexity queries | Vendor partner pages |
| SE Ranking | Cited three times in one Perplexity answer | Third-party study |
Two patterns hold across every sampling so far, and one moved between rows and is reported with both numbers. On the eight commercial questions the first-party share of observed citations went from 43% on 2026-07-16 to 69% on 2026-07-23, so which class of source dominates is itself unstable run to run: on the fixed-price question in entry 3 all eight cited sources were the named vendors' own sites, while the same category question one week earlier cited six third-party listicles and zero vendor pages, and every white-label vendor Perplexity named was cited from its own partner page. On the three data questions the citations go to third-party research instead: Muck Rack, Foundation, arXiv, SE Ranking. And the two engines draw on almost fully disjoint source pools: on the citation-share question, ChatGPT cited Muck Rack, Foundation and arXiv while Perplexity cited five different publications of the same statistic, with zero overlap between the two lists.
What this table does and does not say: it describes who gets cited today, per engine and per question intent, in these dated samples. The vendors cited from their own pages carry authority we have not demonstrated, none of this is evidence that publishing a similar page produces a citation, and our own number on this panel remains 0 of 11.
Nicolò Brignoni
Founder, Deep Ocean · follow the work
Protocol
Every entry in this index follows the same protocol: a fixed, buyer-shaped query panel sampled via the ChatGPT web interface. For each query we record two separate things: the brands named in the answer text, and the source domains the engine cites, whether they favor us or a competitor. The row states what was named, never why it was named or how to change it. Rows are immutable: any panel change is a version bump with a visible break annotation, and retired queries keep their history.
The intended cadence is weekly, and this is the record so far, gaps and all. Entry 1 was sampled 2026-07-08 on panel v2, the 8 buyer questions shown above. The panel then moved to v3 (11 queries: the 8 vendor-recommendation questions plus 3 data-source questions the owned research lane can actually win), so entry 1 and the v3 rows are a versioned break, not a like-for-like continuation. Entry 2 is the first v3 row, sampled 2026-07-16, 0 of 11, and entry 3 is the second, sampled 2026-07-23, 0 of 11. When a scheduled weekly sample is not run on time we show the gap rather than backfill or hide it, and we publish the misses on the same page as the hits.
Known comparability break at entry 1. Entry 1 was sampled in a signed-in web session with memory active. From week 2 the panel is sampled in temporary chat, which carries no memory between runs even though the session is still signed in, so runs are cleaner and comparable to each other. That mode change sits exactly at the start of the series, so we annotate it here: entry 1 and the temp-chat rows that follow are not strictly like-for-like, and we flag it rather than smoothing over it.
Known late run: week of 2026-07-13. The row due that week did not run on the intended schedule. It was completed and published late, on 2026-07-16, once we fixed how the weekly sample gets run. That row is entry 2 above; we mark the slip here rather than let the version-break note above absorb it silently.
Three entries are still three data points. We will not draw a trend line or claim movement in either direction until at least 4 comparable weekly rows exist, and entry 1 sits on the other side of a panel version break from entries 2 and 3, so it is not like-for-like with them. Everything above this line is entries 1 to 3.
This index draws from the same sampling record as our own evidence report. Read engagement #0, the field sample of two absent US software brands, or the AI Readability Reports.
Run this on your own category
This page is our category, sampled on our own panel. The $490 Visibility Sprint runs the same kind of record on yours: your buyers' questions, which brands the answers name instead of you, the sources those answers cite, and crawler access checked per engine, delivered as dated raw records within 48 hours of confirmed scope. It names and sizes the gap. Getting into those sources is work your team executes, and no citation flip is proven yet, ours included.