A page on my own site picked up 7,480 impressions across a ninety day window, sat at an average position of 8.5 on Google’s first page, and earned seven clicks. Click-through rate of 0.09 percent. The textbook reaction is to rewrite the title, tighten the meta description, and wait for the next refresh. That reaction is almost always wrong. Nobody typed those searches. The impressions were logged by Google’s AI Mode and AI Overviews expanding a single user prompt into a fan-out of narrower sub-queries, retrieving my URL as a candidate source for some of those sub-queries, and merging the answer back into the chat. Each retrieval produces a real impression at a real position. None of them produces a click, because no human ever saw a results page. Standard click-through rate analysis cannot tell the difference, and worse, it actively ranks these phantom pages at the top of its “underperformers” list. Below is how to tell which of your pages are being cited by AI rather than read by humans, and what to do once you know.
Why the gap gets misread
The numbers look identical from the Search Console dashboard. A page with high impressions and almost no clicks could be a real underperformer with a weak title, or it could be a page that AI Mode is citing as a source without sending any traffic. The dashboard cannot separate the two. The summary view gives you clicks, impressions, average position, and average click-through rate. It does not give you the distribution of impressions across positions, which is where the difference lives.
The deeper issue is that the standard heuristic for finding “underperforming” pages inverts the signal for AI-cited pages. The heuristic looks for rows where the actual click-through rate falls below the expected click-through rate for the page’s average position. A page that AI Mode is citing with zero clicks has the largest possible gap between expected and actual, because zero divided by anything positive is the maximum. The phantom page rises to the top of your “needs a title rewrite” queue, even though rewriting the title cannot possibly produce more clicks when there were no humans to begin with.
What AI Mode actually does to the index
When a user types a prompt into AI Mode or sees an AI Overview in regular Google, the model does not run that single prompt against the search index. It runs a technique called query fan-out: the model decomposes the prompt into a set of narrower sub-queries, retrieves candidate sources for each one against the real index, ranks the candidates, and merges them into a single response. Each of those sub-queries is a real search against the real index. Each time a page is retrieved as a candidate source, Search Console logs an impression at the position where it was retrieved. The impression is genuine. The position is genuine. The click is structurally impossible.
This is why a title rewrite does not change the outcome on these pages. The page is being cited by the model as part of its own answer. The user is reading the answer in the AI box, not clicking through to the source. If you want to know whether your page is being cited, Search Console does not show that directly. It shows the impressions that the citations produced.
What to actually measure
The cleanest signal is the position-band distribution of your page’s impressions, plus the click-through rate inside each band. A position band is a bucket of average search positions rather than a single number. The simplest split is four bands: top 3, rest of page 1, page 2, and page 3 plus. The bands matter because a single average hides the spread. A page can average position 8 by sitting at position 2 for a handful of searches and 30 for everything else.
For a real human audience, every band should produce some clicks. The top 3 band normally converts at 10 to 40 percent for human searchers. The rest of page 1 converts at 2 to 10 percent. Pages 2 and 3 convert much lower, but they still produce some traffic. If every band produces zero clicks, with a non-trivial number of impressions in the top bands, the page is not being clicked because no human is seeing it.
The opposite signal also matters. If your page produces clicks in the top 3 and in the rest of page 1, but zero clicks on page 2 and page 3, that is normal human behaviour. People click on the first page and rarely go further. That is not a phantom pattern.
How to pull the data without writing a script
You can pull position-band data from the Search Console UI without writing anything. Open the page filter for a specific URL, switch the dimensions to query, set the row limit to a thousand, and export to CSV. Then split the rows into the four position bands and sum the impressions and clicks per band. The numbers you want are the total impressions and the total clicks per band, not the average. Average position within a band still hides spread, but it is a smaller spread than a single page average.
If you want to automate the band split, a fifty-line script against the Search Console API will do it. The script needs three things from Google Cloud: a service account with the Search Console API enabled, a JSON key for that service account, and the service account added as a user on your property with at least read access. The Search Console API exposes a searchAnalytics/query endpoint that returns rows grouped by query, with average position and click count. Bucketing the rows by their position into four bands is straightforward arithmetic on top of the API response. None of this is novel. The interesting question is not how to do it. It is what the bands tell you.
Reading the pattern in your own data
Here is roughly what the band table looks like for an AI-cited page on my site, after splitting the ninety day window by position:
- Top 3: a hundred-ish impressions, zero clicks.
- Rest of page 1: a few thousand impressions, zero clicks.
- Page 2: a hundred-ish impressions, zero clicks.
- Page 3 plus: a smaller number of impressions, zero clicks.
A real human audience never produces zero clicks across every band. A weak title can drop your click-through rate in the top bands, but it cannot push every band to zero, because searchers will still click something. If every band is at zero with thousands of impressions at positions that should always convert, the page is being cited by AI Mode and not being clicked by humans.
The next check is the queries themselves. AI Mode fan-out produces a long tail of narrow, oddly specific queries that real searchers rarely type. If you export the top queries by impression and they look like fragments of your own article rather than real questions, you are looking at AI fan-out traffic. If they look like questions a person would type, you are looking at real search traffic, and your title really does need work.
What to do with the answer
Once you know which pages are AI-cited, the right action depends on which audience you wanted.
If you wanted organic human traffic, you do not have it, and rewriting the title will not produce it. The traffic is being absorbed by AI Mode at the citation step. The honest action is to accept that this page is being used as source material for AI answers, and decide whether that is the role you wanted the page to play. If not, the fix is to make the page less citable, which is a content decision, not an SEO decision.
If you wanted AI citations, you already have them. Search Console just made them visible. The standard AI citation metrics are emerging, but the cleanest signal right now is the position-band pattern described above. A page with non-zero impressions in the top bands and zero clicks is being cited. That is a successful outcome if your goal was citation.
Either way, stop rewriting the title. The title is not the problem.
Trade-offs
Knowing the band-level distribution is more useful than the page-level average, but it is not a complete answer.
- The data lags by two days. Decisions you make today are looking at a window that ended on Tuesday. If you are chasing a specific incident, that lag is the time you cannot recover.
- The position-band thresholds are opinionated. I chose top 3 plus zero clicks as the most reliable phantom signal, because that band converts highest for human searchers. A page with very few impressions in any band cannot be classified confidently. The pattern works above a few thousand impressions. Below that, you are guessing.
- Average position within a band is still an average. A page can be at position 5 for one query and position 9 for another, both inside the rest of page 1 band. The band collapses them. If you need finer resolution, group by individual position or by query.
- A phantom page is not always a bad page. If your goal is to be cited by AI Mode, this is exactly the traffic pattern you want. The analysis tells you what is in your data. It does not tell you what to do about it.
If you only do one diagnostic, pull the band-level numbers for your top-3 URL by impression count. The control is your best-performing article, which should show real clicks in every band. The contrast between the two tells you which pages are real and which are AI-only.
What to take away
Three short notes I would give to anyone staring at a Search Console report full of zero-click pages.
- Zero clicks at every position band, with non-trivial impressions in the top bands, is an AI citation pattern, not a title problem. Stop rewriting titles on phantom pages.
- The position-band distribution is the signal the dashboard hides. Pull the query-level data, split it into bands, and look at the click count per band. The average lies. The bands do not.
- Decide which audience you wanted before you start measuring. AI citations are a different metric from search clicks. Standard Search Console tooling conflates them. Once you separate them, the right action becomes obvious.