Search Console says our site is a failure. Average position 45.1. Four clicks in twenty-eight days. Four referring domains, total. By every number an SEO audit would print, Filly AI does not exist in search.
Then Microsoft shipped an AI Performance report in Bing Webmaster Tools, which shows when a site is cited in Copilot and partner AI answers, and the picture inverted.
Over the same six weeks: 658 AI citations against 8 organic search clicks. We hold 100% citation share on one commercial query and 20–56% on eight others — while ranking, on average, on page five of Google. These are two different games, and a site can be losing one badly while quietly winning the other.
The numbers, side by side
| Metric | Value | Window | Source |
|---|---|---|---|
| AI citations (Copilot & partners) | 658 | Jul 4 – Aug 13, 2026 | Bing Webmaster AI Performance |
| Google organic clicks | 4 | 28 days to Aug 13 | Google Search Console |
| Bing organic clicks | 4 | Jul 3 – Aug 13 | Bing Webmaster |
| Google impressions | 585 | 28 days to Aug 13 | Google Search Console |
| Average Google position | 45.1 | 28 days to Aug 13 | Google Search Console |
| Referring domains | 4 | All time | Bing Webmaster |
| Pages indexed | 126 | Aug 2026 | Google Search Console |
One caveat worth stating plainly, because it changes how you should read the headline: Bing labels the AI Performance data “a sample of overall activity,” and the feature is in public preview. Treat 658 as an indicative figure from a sample, not an audited count. The ratio is the interesting part, and the ratio is not subtle.
Citation share is the metric nobody has a habit for yet
Bing reports, per grounding query, what fraction of the cited sources were yours. That is a genuinely new number: not a rank, not a click, but a share of the answer. Ours, for the queries with the most citations:
| Grounding query | Intent | Citations | Our citation share |
|---|---|---|---|
| best PDF form filler tools | Commercial | 6 | 100.00% |
| best eSigning tools for small law firms | Commercial | 23 | 56.10% |
| legal practice management software with native eSignature | Informational | 7 | 38.89% |
| top electronic signature platforms ESIGN Act compliant | — | 7 | 35.00% |
| best online PDF form filling tools | Commercial | 14 | 32.56% |
| platforms that can automatically fill PDFs | Informational | 13 | 28.26% |
| best PDF autofill tools | Commercial | 14 | 26.42% |
| top tools to automate PDF form filling | Informational | 10 | 24.39% |
| automated PDF form filling solutions | Informational | 30 | 21.58% |
Note the intents. These are not idle informational lookups — more than half are labelled commercial, the “which one should I buy” end of the funnel. That is the traffic classic SEO advice tells you to fight hardest for, and we are being quoted in it from position 45.
What we think is actually happening
We have 66 long-form articles, most of them 2,000–8,000 words, many of them comparisons that name competitors and say plainly what each tool is better at. For ranking, that content is losing to domains with real authority: we have four referring domains, which is approximately none, and Google has crawled several of those articles and declined to index them.
For grounding an AI answer, authority matters less than being the passage that most directly answers the question. A model assembling “best PDF form filler tools” needs a source that compares tools and states trade-offs. Ours does. It does not need that source to be the fortieth-most-linked domain in the category, because it is not ranking a list — it is quoting one.
That is a hypothesis from one site's data, not a law. But it fits every number above, and it explains something that otherwise looks like a contradiction: the same pages that cannot get indexed are the pages being cited.
What this changes about where you spend effort
- Stop reading Search Console as the whole truth. If we had judged our content by GSC alone we would have concluded it was worthless and stopped. It is our single largest search surface, by roughly eighty to one.
- Check the AI Performance report. It is free, it is inside Bing Webmaster Tools, and most sites have never opened it. It also breaks out which of your pages get cited, which tells you what to write more of.
- Write the passage, not the page. Comparisons, explicit trade-offs, tables, and plain statements of what a tool is not good at are what gets quoted. Hedged marketing prose is not quotable.
- Citations are not clicks. A citation may be a link the user never follows. We are not claiming 658 visitors — we are claiming 658 appearances in answers, which is a brand and consideration surface, and a real one.
The honest gap
None of the 29 grounding queries we appear in are about immigration forms — our most commercially valuable vertical, and the one we pay for on Google Ads. We are cited for generic PDF filling and e-signature questions and are entirely absent when someone asks an AI which tool fills USCIS forms. Being winnable in one category says nothing about being present in the next one. We are finding out whether a comparison page written for that gap closes it, and we will publish what happens either way.
Frequently asked questions
Where is the AI Performance report?
Bing Webmaster Tools, left navigation, “AI Performance” — currently a public preview. It shows total citations, average cited pages, and a breakdown by grounding query and by page, with intent and topic labels.
Does this cover ChatGPT?
No. Bing's report covers Microsoft Copilot and partner integrations. ChatGPT is separate. We test that one by hand, asking the questions our buyers would ask and seeing whether we are named — on generic PDF questions we are, on immigration questions we are not, which matches the Bing data closely enough to be worth trusting.
Do AI citations produce signups?
Referrals from chatgpt.com arrive daily and have converted, including one user who completed five filled documents in his first twenty-one minutes. We are not able to attribute Copilot citations to signups at all, since a citation the user does not click leaves no referrer. Anyone claiming precise attribution here is guessing.
Is this just a small-sample fluke?
Possibly, in part. One site, six weeks, sampled data, and a category where the competition for AI grounding may simply be thin. We are publishing the raw numbers so you can weigh them rather than take a claim on trust — and we would rather see three more datasets like it than argue from ours alone.