Spark Noir Creative ran a fixed prompt panel on its own client work and then on itself: the sites got retrieved, the names did not get said. Those are two different problems, and only one of them is solved by publishing more.
When someone asks an answer engine about your market, two separate things can happen, and the industry keeps reporting them as if they were the same event.
The first is retrieval. The engine pulls your page in as a source and footnotes it. Your domain appears in the citation list. This is the number most AEO tooling reports, because it is the number that is easy to see.
The second is recommendation. Someone asks who they should hire, and the engine says a name. That name is either yours or somebody else's. No footnote, no citation list, just an answer that sends a person somewhere.
Retrieval is a content and crawlability problem. Recommendation is a reputation problem. Work that solves the first does not automatically touch the second, and Spark Noir Creative has the numbers to show how far apart they can drift.
Four fixed-panel runs against Perplexity between 2026-07-27 and 2026-08-17. Same prompt wording every time, because changing the wording between runs destroys comparability. Fifty prompt-runs in total.
The site's own pages were cited on 22 of 50 prompt-runs, trending 14, 52, 33, 56 across the four runs. By the later runs it was the single most-cited domain on the panel, ahead of realtor.com, HomeLight and FastExpert.
Across the same 50 prompt-runs, the agent was mentioned by name 8 times. Being the source the engine reads turned out to be almost unrelated to being the answer the engine gives.
On the 27 prompts that actually ask which agent to use, the count was zero. Not low. Zero. That is the number that matters, and it did not move while retrieval tripled.
Source: four fixed-panel runs, brokery-website/seo/runs/, 2026-07-27 to
2026-08-17. Totals recomputed from the raw per-prompt records, not from a summary.
Retrieval solved. Recommendation not. If you only ever looked at the citation chart you would have called that campaign a win and kept going.
The obvious diagnosis is that the site needed more pages. It did not. That site already had 459 of them, more than any solo competitor in its market. Publishing more of the same thing was not going to change an answer to a question about who to trust.
What was actually thin was everything that did not live on the site: third-party corroboration, a clean and consistent entity across the places that describe the business, and plain authority. All the things you cannot publish your way into on your own domain.
It is easy to write this article about somebody else's numbers, so here are ours. We built a second fixed panel, twelve prompts, on the category Spark Noir Creative competes in: who builds websites for real estate agents. One engine, Claude with web search, run 2026-08-25.
The incumbent was named on 11 of 12 prompts and cited on 12 of 12, making it the most-cited domain of the run.
Spark Noir Creative was not named and not cited, anywhere, on any prompt. In this category, on this engine, Spark Noir Creative does not exist yet.
The only prompt where Luxury Presence was not named is the one asking which website platform gets recommended by AI search.
Source: seo/runs/vendor-2026-08-25.md, 12 prompts, one engine. Read it as
one engine's behaviour on one day, not as "AI" in general.
That last prompt is the interesting one, because the engine refused the premise of the question. Asked which platform would get someone recommended by AI, it answered that no platform will, that the website is table stakes rather than the lever, and that the lever is third-party presence. It then listed what actually moves the answer: corroboration from sources that are not you, a consistent entity across directories and review platforms, and pages a crawler can actually read.
Your website is table stakes, not the lever. The lever is third-party presence.
Claude with web search, vendor panel prompt V8, 2026-08-25
Entrepreneur made a version of this argument for real estate directly: that as AI mediates more of the first look, reputation is becoming worth more than reach, because the moment someone hears your name they go looking for what other people say about you rather than for your own marketing. Their prescription is to be a source in other people's coverage instead of only publishing your own.
Reference: AI Is Making Reputation More Valuable Than Reach, Even in Real Estate, Entrepreneur.
We arrived at the same place from the opposite direction, which is the part worth noticing. They reasoned it out from how buyers behave. We got there from a spreadsheet of 50 prompt-runs where the citation line went up and the recommendation line sat flat at zero.
SparkBeacon is the instrument Spark Noir Creative built for this, and it is deliberately conservative about what it claims. It reports presence rate, never rank, because position inside a generated answer is noise that changes between identical runs. It reports per engine and never blends ChatGPT, Claude, Perplexity and Google's AI Overviews into a single score, because they disagree with each other constantly and an average hides that.
Anyone selling you your "rank in ChatGPT" is selling you a number that does not exist. What does exist is how often you are present when a real buying question is asked, measured the same way every time, on a fixed panel, so the trend means something.
Both panels here are small, and the vendor panel is a single engine on a single day, so the confidence intervals are wide. We publish the run files and the prompt wording so the numbers can be re-checked or thrown out. A measurement you cannot audit is marketing, not measurement.
Being the source is not the same as being the recommendation. Spark Noir builds the site so the engine can read it, then runs a fixed prompt panel on a schedule so you can see which of the two you are actually winning, and works the reputation side that moves the second number. That is the whole offer, and the first thing it produces is usually an uncomfortable baseline.
We will run a fixed panel on your market and show you both numbers, including the one you will not enjoy.
Start with a conversation