We Wrote Four Takes on AI Overview Ads. One Query Got 543 Impressions and One Click

One click from 543 impressions. That is what Search Console shows for our query ai overview ads: an average position of 50 after we put at least four near-identical takes on the subject in our backlog. I have told clients for years not to split one question across several URLs. We did it anyway.

This is a postmortem of our own mistake, not a theory about somebody else’s site. The first sign was a weak search result. The more uncomfortable sign came from our edge logs: the crawled AI Overviews ads page had 381 bot hits, only 18 verified, and zero live answers. A dated Google Ads changelog on the same site appeared in 355 live answers. We kept producing new pages when we needed to maintain one.

The search figures come from our Search Console query view; the fetch and answer counts come from our own edge logs. They are not one conversion funnel. The query figures do not tell us how impressions were divided among the overlapping pages, and a bot hit is not the same measure as a live answer. Nor does this comparison prove that consolidation alone would have produced citations. It does show where our attention went, what the observed result was, and why another draft was the wrong next move.

The setup: a placement advertisers could not isolate

In October 2025, Google confirmed that ads could appear above, below or within AI Overviews. An ad inside an Overview had to be relevant to both the query and the Overview content. Advertisers could not target that placement directly or opt out of it. Eligible Search, Shopping and Performance Max campaigns could enter the auction.

That gave advertisers a real question: what, if anything, should they change? The answer needed care because eligibility was not the same thing as a separate set of placement controls. Google pointed to AI-powered matching, including broad match, AI Max, Performance Max and Dynamic Search Ads. The queries behind an Overview could fan out into related questions an advertiser never typed into a keyword list. I understood that distinction in the ads account. On the blog, we behaved as though a fresh headline was a fresh answer.

We had a reason to publish quickly. The placement was new, clients wanted a plain-English explanation, and the rules were worth covering. So we published an October 2025 guide. Then the subject came around again, and we answered substantially the same question in another draft. Across the backlog, we ended up with at least four near-identical pieces, including that guide and later takes that approached the ranking question from slightly different angles.

None of this required a grand content strategy to go wrong. One news cycle, one new title, one more introduction promising to explain what advertisers needed to know. Repeat. The question stayed put while the URLs multiplied.

The first sign: search had demand, but our answer did not land

The Search Console result should have stopped the next draft before it started. The query ai overview ads drew 543 impressions and one click at an average position of 50. That does not tell me every reason the page ranked poorly. It does tell me people were searching, Google was showing our result, and the result was not winning much attention.

At that point, the useful question was not, “What else can we publish?” It was, “Which page are we asking Google to treat as the answer?” We had no clear response. An October guide remained in the mix while later pieces revisited the same subject. Someone arriving through any of them was trying to resolve much the same issue: how ads in AI Overviews work and what an advertiser can actually do about them.

I used to tell clients that one clear page beats three pretty good pages when all three serve the same searcher. I was wrong to act as though our own blog was an exception. The warning was not a mysterious algorithm signal. It was the overlap visible in our backlog, next to a query result nobody would put in a performance deck.

When several of your pages answer the same searcher, count them before commissioning the next one. We did not.

The wrong call: we refreshed the headline, not the answer

The first wrong call was speed. Each time the topic returned, we treated another article as the way to stay current. The better editorial decision was to revisit the existing answer as Google clarified the placement. Instead, we gave the same underlying question another URL, another introduction and another chance to compete with our own work.

That is the problem usually called content cannibalization: pages that chase the same intent rather than building a distinct answer. The label is less important than the practical effect. A reader has to choose among similar pages, an editor has to keep several versions accurate, and links and crawl attention have more than one place to land. None of those choices makes the answer clearer.

The second wrong call was framing. Too many of the drafts promised some version of what advertisers need to know. That sounds comprehensive until you ask which question a paragraph answers on its own. A broad overview can be useful to a person reading from the top. It is less useful when a system needs a short, self-contained answer block for a specific question buried inside the topic.

That distinction mattered here. Google’s query fan-out means an Overview may draw on pages addressing related questions, not just pages ranking for the exact wording a searcher entered. An Ahrefs study of 863K SERPs found that 37.9% of cited pages also ranked in the top 10. That study does not diagnose our URLs, but it explains why “write another broad article and wait for the head term” was a thin plan. We needed sections that made individual answers easy to find and understand. We kept writing introductions that tried to cover the whole subject again.

I would have told a client to stop at this point. We did not stop ourselves. A new title is not a new question.

The damage: visits without a live answer

The edge logs made the failure harder to shrug off. The crawled AI Overviews ads page recorded 381 bot hits. Only 18 were verified, and we recorded zero live answers for it. Total hits looked like activity until we separated verified attention from background traffic. After that, there was not much to celebrate.

We already knew how to make that distinction. We have published a teardown of which pages AI crawlers actually fetch. Yet we let our own cluster sit without making the fetch and answer check the editorial decision point. I can forgive a page that does not immediately earn a citation. I am less generous about publishing another version of it before looking at the logs we already have.

There was also a limit to what the articles could responsibly promise. Advertisers have no isolated CPC or ROAS for ads shown inside AI Overviews. If a draft suggested that an advertiser could verify placement-specific performance tactics in-platform, it went further than the reporting allowed. More articles did not solve that constraint. They gave us more places to keep the explanation straight.

Now compare what we could actually observe on our own site. The Google Ads updates changelog at /post/google-ads-updates-2026 logged 576 verified fetches and 355 live answers. Those figures do not establish that its format caused the difference, or that the ads pages would have matched it if merged. The pages cover different material. But the changelog offered something our cluster did not: dated entries on a maintained URL, rather than another article each time the subject moved.

Each entry gave a reader a way to see what was current. Each update gave us a reason to improve the same page instead of asking a new one to start from scratch. We had seen a similar contrast in our logs before: one dated changelog drew 182 live-answer fetches while three category pages drew none. That earlier observation was not a guarantee either. It was a reason to take the maintained page seriously as an editorial model.

What the deck calls content velocity, I call four people digging the same hole. The changelog is one hole, dug deep, with dates on the wall so you can see where you are. The logs told us which page was being used; they did not give us permission to claim a result we had not achieved.

The fix we should have made before the fourth draft

The decision that would have prevented the overlap was simple: choose one canonical answer to the advertiser’s question and keep it current. Not one page for every Google Ads update. One page for this recurring question about ads in AI Overviews, with distinct, dated sections as the facts changed.

First, we would choose a survivor from the overlapping pages. We would compare what each one answers, what is still accurate and which URL search engines or verified assistants already fetch. The oldest page would not win merely because it was first. Nor would the newest win because its prose reads better. Good sentences can move; the point is to stop making a reader or a crawler choose among several versions of the same answer.

Next, we would move the passages worth keeping into that survivor and 301 redirect genuinely redundant URLs to it. A canonical tag signals a preferred URL while leaving the other pages available. For pages with no separate job to do, a redirect makes the choice clearer. This is a proposed repair, not a result I can claim we have already measured. Search Console and the edge logs would have to tell us what changed afterward.

The surviving page would need to earn its keep. I would give each distinct question a heading and a short answer that states the claim up front, then add the detail an advertiser needs to judge it. Date meaningful updates. When Google clarifies eligibility or reporting, revise that answer rather than publishing another general guide. A 40-to-60-word answer block is a useful editing constraint, not a promise that an assistant will quote it.

There is an important boundary here. I would not redirect two pages merely because they mention AI Overviews. If each serves a different buyer question, each deserves its own URL. Our problem was the near-identical answer repeated across the backlog. Consolidate the answer, not the entire subject.

After the repair, I would watch the same measures that exposed the mistake: Search Console for where impressions and clicks land, verified fetches for whether assistants return, and live answers for whether they use the page. I would keep those measures separate. An impression is not a click. A hit is not a verified fetch. A verified fetch is not a citation. Collapsing them into a single story is how a busy chart starts doing the thinking for you.

I learned the equivalent lesson managing paid accounts. When a structure is fragmented, adding more of the same does not make it coherent. You identify what each part is for, consolidate what is doing the same job, and then give the survivor your attention. That is the judgment we should have applied to the blog before the next news cycle arrived.

The numbers do not tell me that one maintained AI Overviews ads page would have reached the changelog’s 355 live answers. They tell me that our repeated drafts left us with one click from 543 impressions on the query, while the maintained page gave assistants something they used. That is enough to decide against a fifth take.

One question gets one page, merged and kept current. Next time the topic spikes, I update the survivor.