Who got cited instead of you?
You give it a question. It forms the retrieval pool, runs the assistant, and reports which sources were relied on — plus what those sources carried that yours did not.
the live tool opens with the free tier · the worked example below is real data from our published experiment
Fourteen pages were fetched. Between two and five were cited.
This is the part most tools hide. You are not competing against the whole internet — you are competing against a short list that a retrieval system already decided was worth reading. Eight of these fourteen pages were read and then thrown away every single time. If you are not on the list at all, nothing on your page matters.
# query best ai ppt maker # retrieval pool 14 unique sources # cited, run by run 2 · 5 · 3 · 2 · 5 # run-to-run overlap 0.472 <- the least stable of our five # cited in every usable run 1 / 14 (ppt.ai, 1,861 words) # never cited, in any run 8 / 14
Only one page out of fourteen was cited every single time, and it is not the longest one. The rest of the citations moved around between runs. That movement is the story, and no tool in this category reports it.
Letting the model read the page helped some and hurt others
Three pages that were never cited when the model saw only titles and snippets became cited once it could read the page — aipowerpointmaker.com went from 0 of 5 to 3 of 5. Four pages went the other way, cited at least once on snippets and never on the page: dokieai.com went from 4 of 5 to 0 of 5. This condition is not a rising tide. And the longest page in the pool — 6,105 words, ranked first — was never cited under either condition.
# gained — 0 on snippets, cited on page aipowerpointmaker.com 0/5 -> 3/5 m.aippt.com 0/5 -> 1/5 presentationgpt.com 0/5 -> 1/5 # lost — cited on snippets, 0 on page dokieai.com 4/5 -> 0/5 microsoft.com 3/5 -> 0/5 wps.com (blog post) 1/5 -> 0/5 ppt.thirdme.com 1/5 -> 0/5 # longest page, ranked first, cited neither way aipresentationmakers.com 6,105 words 0/5 -> 0/5
Both conditions ran the same five queries against the same pools with the same model. The only variable was whether the model could read the page body. That the effect runs both ways is exactly why we report counts instead of a score.
Some questions are more stable than others
Run-to-run overlap across our five queries runs from 0.472 to 0.700. The least stable — best ai ppt maker — had between two and five of its fourteen sources cited from one run to the next. The most stable — robots.txt allow ai crawler gptbot — was answered from the same two or three sources every time. Stability is a property of the question, not noise you can average away, and it tells you where a new page still has room.
# most stable — overlap 0.700 robots.txt allow ai crawler gptbot the same two or three sources every run # least stable — overlap 0.472 best ai ppt maker 2 to 5 of 14 cited, and the set churns # we publish the run count, not a rank
We will not give you a visibility score. Three tools measuring the same brand against the same forty prompts have been observed to report 22, 19 and 24 hits. What we report is a decision: which questions are worth your time.
What the live tool will do, and when it opens
The live version takes your question and your domain, forms a fresh retrieval pool, runs the assistant six times, and reports the cited set, the overlap, and a comparison against the pages that beat you. It is free once a day.
It is not open yet, and the reason is arithmetic rather than engineering. Every run costs a search query and a model call, and the cost scales with how many people use it. We are not going to open it, watch it get hammered, and then close it — so it opens with the free tier, with a daily limit, and it stays open.
In the meantime the two tools that cost us nothing to run are open and unlimited. The readability check and the fix generator both run in your browser, and between them they cover the two checks we have actually measured against citation rates.
Questions about the citation test
Can I run it now?
No. The tool is not open yet, and the button on this page is disabled rather than pretending otherwise. What is on this page is the measurement we have already run, with the method behind it.
Why will there be a daily limit?
Because every run costs model calls. The readability check is free and unlimited because it only reads one page and calls no model; this one asks a model the same question repeatedly, and that has a price per run.
What does one run actually tell me?
Which sources were cited for one query in one run, and which were reachable but skipped. It is a snapshot, not a score: run-to-run overlap on the same query ranged from 0.472 to 0.700 in our own experiment.
Why does the same query give different answers?
Because the cited set is not fixed. Of our five queries, the most stable had a run-to-run overlap of 0.700 and the least stable 0.472. One run cannot tell you whether you have a problem or a bad draw.
Where do the numbers on this page come from?
From one experiment: five real queries, 48 unique sources returned by a web search, and a fixed model asked the same question six times per query. 28 of 30 runs were usable. Every figure can be recomputed from the method page.