The AI Search Dominance Engine: 30 Prompts, 7 Waves, Measured in Public

What the series shows

Seven dated rows, 2026-09-05 to 2026-10-03 — 6 of them scored runs and the first a sampled baseline. Each scored wave is the sum of 30prompts scored 0–3, so 90 is the ceiling. The band is flat and it is close to the floor.

Cite as: The AI Search Dominance Engine, Digital Services Program, wave of 2026-10-03. Free to quote and reuse with attribution (CC BY 4.0). Download every wave and observation as CSV.

Six things the data says

1. A month of correct on-page work moved the retrieval layer by nothing

Across every wave, digitalservicesprogram.com has never once appeared for a non-brand prompt. Not at the peak of 7, not at 3. Every point the series has ever scored came from a third-party property DSP controls, or from its own brand name.

Over the same weeks the site shipped and verified: one Organization and one Person entity with shared identifiers across every page, answer-first definition blocks reused verbatim as structured-data descriptions, FAQ arrays rendering the same text visibly and machine-readably, an explicit AI-crawler allowlist, per-URL sitemap dates wired to a content-change constant, and a blog consolidation with redirects. None of that is wrong. The finding is that it is a floor, not a lever — a site with no third-party corroboration does not get retrieved however clean its markup is.

2. On buyer queries the results are not competitors, they are directories

On Best AI training companies in Islamabad(5 October) the entire top three was DesignRush, GoodFirms and TechBehemoths, with no company's own website until fourth. On a related phrasing measured on 3 October outside the fixed prompt set — which company builds AI employees in Pakistan — nine of nine results were directories and no company's own site appeared at all. No Pakistani AI company in this measurement ranks for its own category on the strength of its own site.

The practical consequence: for this class of query the achievable goal is not outranking the directory. It is being the top-listed company inside it.

3. A third-party profile spikes, then decays if nobody maintains it

github.com/DSPSardar was the #1 result for Who teaches AI agents in Islamabad?through seven consecutive verifications in September. By 3 October it was gone from that result set, and gone again on the 4th and 5th. The profile's text had not been edited since before DSP changed its own positioning on 15 September — so the highest-retrieval property the company owned was publishing a description the company itself had stopped using.

A directory listing on the same brand query behaved differently: it took #1 within 48 hours of being claimed, lost #1 within about a day, and then held second place on both 4 and 5 October. The pattern that fits both is a freshness spike settling to whatever corroboration supports it. One profile is a spike; a set of mutually-corroborating profiles is a floor; an unmaintained profile decays.

4. The brand query is lost to name collisions, five distinct kinds of them

What is Digital Services Program (DSP) in Pakistan?is the prompt DSP should win most easily. It scores 1. On 5 October the set ran: a differently-owned company with a near-identical name at #1, a directory profile at #2, an unrelated encyclopedia article at #3, a generic industry-term explainer at #4, and digitalservicesprogram.com at #5 — its best position yet on that phrasing, and still a 1, because no owned property is top three.

Five collision classes are now measured on that single query:

1. A differently-owned company with a near-identical name
2. DSP = Deputy Superintendent of Police, the dominant Pakistani reading of the acronym, which took two of the ten slots
3. Government “digital services programme” initiatives in other countries
4. “Digital service provider” as a generic industry term
5. DSPy and demand-side platforms in the wider technology sense

A name that collides five ways is not one an engine can resolve by consensus. Disambiguation has to be published as content — the identifiers that are uniquely yours, stated where a machine and a human can both read them. DSP's own answer to that is on its about page.

5. The paid placement we refused is ranking

A Pakistani technology outlet offers a bylined “best AI academy” piece as a commercial placement. DSP refused it, on the grounds that paying for a superlative is buying a citation rather than earning one. On 3 October that article ranked third on What are the best AI education platforms in Pakistan— a prompt DSP scores 0 on.

That is the honest cost of the refusal and it belongs in the data rather than out of it. The defensible position is not that paid placements fail in the short run. It is that a citation you bought is a citation anyone can later discover you bought.

6. A directory's category page gated on the service tag, not on reputation

Tested with a control group rather than guessed at, on 23–24 September and outside the fixed prompt set. On one directory's artificial-intelligence category page for Pakistan, four listed companies were sampled: all four had no reviews at all, and one listed AI as a fifth of its business behind a much larger mobile practice. DSP's listing, whose service was tagged “AI Development” rather than “Artificial Intelligence”, did not appear on it.

The gate was the tag string, not reputation. The generalisable practice is to sample the marginal listings on a category page and compare their fields with yours before changing anything.

How a wave is scored

• One search per prompt, per wave. No re-runs and no best-of — a second look at a disappointing result is how a series stops being a measurement.
• Scored by three independent agents, ten prompts each, none of them told which company the run belongs to or what result is hoped for. Priming the scorer is the easiest way to fake a dataset like this, so it is the thing most deliberately prevented.
• The #1 result is recorded for every prompt, which is what lets a stranger re-check a row months later.
• A DSP property means digitalservicesprogram.com or any profile DSP controls — its GitHub, its directory listings, its social pages. Third-party pages that merely mention DSP do not count.
• Full waves run weekly. Five-prompt spot-checks between waves are recorded as observations and never mixed into the /90 series.

All 30 prompts

Fixed on 2026-09-05and never edited since — including the four now known to be read by engines as a different question than intended, which are kept in the set, scored anyway, and flagged below. Changing a prompt because it embarrasses you is exactly what would stop the series being evidence.

Recorded observations

Per-prompt rows were kept from 2026-10-03 onward, plus the individual observations recorded before that. The #1 result is included on every row, which is what makes these re-checkable rather than merely reportable.

What this does and does not show

• The search tool used resolves to a US locale. A Pakistan-locale run would plausibly score differently on the Islamabad and Pakistan categories, and this is the largest known weakness in the series.
• One sample per cell. Result sets are volatile, so a single-wave move is noise and only a trend across waves is signal.
• This measures the retrieval layer — the live index that ChatGPT Search, Perplexity, Gemini with grounding, Google AI Overviews and Claude web search all draw from before they synthesise. It is a proxy for AI-search visibility, not a measurement of any one engine’s rendered answer, because those vary by account, session and model version and cannot be used as a time series by anyone who does not control those variables.
• Per-prompt rows were retained from 2026-10-03 onward. Earlier waves are published as dated totals plus the individual observations recorded at the time.
• The 5 September figure is an approximate baseline from sampling, not a scored run, and is marked as such.
• This dataset is published by the company it measures. The protections against that are the fixed prompt set, the unprimed scorers, the recorded #1 results — and findings that are mostly unflattering.

What it can tell you is how one small company's retrievability behaved, week by week, while it did the on-page work everyone recommends — and which of its own actions moved the number and which did not. As far as we can find, no comparable dated series is published by anyone. We would rather publish it with its weaknesses stated than not publish it.

Who ran it, and what they get out of it

Sardar Ghaffar, founder of Digital Services Program, an AI Employee company in Islamabad that builds AI Employees for businesses and teaches people to build them. So the interest is plain: DSP sells both, and a page demonstrating rigour about search is a page that flatters DSP.

Which is the reason the prompt set was fixed before the first measurement, the scorers are not told what answer is wanted, the #1 results are recorded, and the strongest finding in the whole series is that a month of DSP's own work achieved nothing measurable. Check the method against the conclusions, and check the conclusions against the CSV.

Corrections and contradicting measurements are welcome and will be published on this page. The next wave is scored weekly.