7 AI Overview Myths I Had to Unlearn as a Google Ads Buyer

The most expensive mistake I made about Google AI Overviews was assuming I could buy my way into a citation. I spent ten years learning to think in auctions: bid higher, write tighter ad copy, mine search terms at 1am, watch CPA come down. That brain is useful for ads. It is lousy at explaining why an AI Overview quotes one page and passes over another. Citations are not ad slots. They depend on whether a page gives the answer something usable to cite, not on what its owner paid to appear nearby.

Myth 1: “Running ads in AI Overviews gets you cited in them”

I believed this one first because it looks like every other Google placement I ever bought. You pay, your ad shows up near an AI answer, and your logo is on the screen. Surely that helps the answer choose you as a source.

That is auction thinking applied to a different job. Paid placement does not buy source selection. The ad earns its place through the paid system; a citation depends on what the answer can draw from the page. Showing up beside an answer is not the same as giving it a passage worth quoting.

The available numbers here come from AI Mode, not AI Overviews, so I would not treat them as an AI Overview citation rate. They do show how little overlap an advertiser should assume. Across 50,032 commercial keywords tested, AI Mode showed a paid text ad on 29.45% of queries, but only 11.53% of advertiser domains were also cited for the same keyword. At the exact URL level, the overlap fell to 1.95%. Only 2.32% of advertised URLs ranked organically for the keyword they bid on, and around 85% had no organic footprint at all.

Buy an ad if the ad can pay for itself. Do not put citation credit in its business case. If the page beneath it says little more than “innovative solutions,” the auction cannot write the missing answer for you.

Myth 2: “A monitoring tool will get you featured”

This is the Quality Score dashboard fallacy with new paint. I used to stare at Quality Score columns as if looking harder would move the number. They were a readout, not a lever. An AI visibility tracker can show that you are absent from an answer. It cannot, by showing you the absence again tomorrow, make your page worth citing.

The distinction matters because these answers move. One oft-cited test found the same brand list in the same order in only about 1 in 1,000 runs. A single screenshot can make a normal change look like a triumph or a crisis. Meanwhile, entry pricing can run $20–$300 a month before any execution work happens. A dashboard is an expensive place to admire a problem if nobody is assigned to fix the copy.

I still want a record of which queries show a cited link, which show a passing brand mention, and which show neither. Then I want someone to read the answer, inspect the competing passage, and change the page. Those are different tasks. If you are comparing tools that help get a business featured, ask what happens after the chart identifies a gap.

Track enough to choose the next edit. Do not mistake the tracking for the edit.

Myth 3: “If you rank top 3, you get the citation”

I used to tell clients to chase the top position and then work on the snippet. That instinct made sense when the visible prize looked tied to a conventional ranking. It breaks down when an AI answer needs several pieces of an answer and can draw them from different pages.

An Ahrefs analysis of 863,000 results found only about 38% of citations came from top-10 organic results, down from about 76% in July 2025. That does not make ranking useless. It means a high position alone is a poor explanation for why one page gets quoted and another does not.

Think about the work the answer has to do. A buyer asks a broad question; the answer may need a price, a limitation, a comparison, and a next step. Your top-ranking category page might cover all four in polished language without answering any of them plainly. A lower-ranking page might put a direct answer to the price question in one short paragraph. For that part of the response, the second page gives the system more to work with.

Ranking and citing are different games with overlapping equipment. Strong pages can help in both, but the passage still has to earn its place. I go deeper on that split in SEO vs. AEO: what gets a page cited. For now, the useful move is simple: find the question your page nearly answers, then make the answer unmistakable near the top. Your rank tracker cannot do that for you.

Myth 4: “Schema markup is the switch”

This belief survives because schema feels like a PPC ad asset: add the right extension, unlock more real estate. It is tidy, technical, and much easier to assign than asking someone to rewrite an evasive paragraph. I understand the appeal.

But there is no citation toggle hiding in JSON-LD. Google’s generative-AI guidance says structured data is not required for generative AI search and there is no special schema.org markup to add. In one matched test, 1,885 pages that added JSON-LD saw no meaningful citation gain in AI Mode or ChatGPT and a 4.6% relative decline in AI Overviews. A separate live-fetch test found ChatGPT, Claude, Perplexity, Gemini, and AI Mode did not extract information that existed only in JSON-LD.

I keep markup accurate. It still has a job in helping search systems understand a page. I just refuse to make it the only place a useful fact lives. If the answer depends on what a service includes, how long it takes, or who should avoid it, I put that in visible copy, in sentences a person can understand without opening the page source.

Here is the test I use: hide the code and read the page. If the answer disappears, the page is not ready. Keep the markup; write the quote.

Myth 5: “AI Overviews cite the biggest brand in the category”

If this were true, my job would be easy. Whoever spends the most on brand search would own the answer box. It is a comforting theory for the biggest advertiser and a discouraging one for everyone else. It also confuses being well known with answering a particular question well.

In a six-week study across five engines, 60% of AI-cited URLs did not rank in the top 20 on Google or Bing for the same query. The study also found that 78–85% of cited domains were unique to one engine. Those findings are not an AI Overview-only scorecard. They do make it hard to argue that every answer simply crowns the same familiar domain.

A buyer asking about a specific installation cost does not need another paragraph about a national brand’s commitment to excellence. A niche installer with a clear pricing page and a photo of the actual work may give the answer a more useful passage than a big brand’s vague “solutions” page. The smaller business does not win by pretending to be bigger. It wins by saying something specific the bigger page left unsaid.

Brand recognition may matter to the person deciding whether to click or buy. It does not relieve the page of the work of answering the question. Write the detail your competitor left out: turnaround time, what the price includes, or what can go wrong if someone skips a step. If you cannot point to the sentence an answer could use, a bigger logo will not supply one.

Myth 6: “One great pillar page covers your whole industry”

PPC brains love consolidation. One tight ad group, one landing page, fewer moving parts. I ran accounts that way for years. The temptation in organic content is familiar: build one enormous page, cover every topic in the category, and call it done.

An AI answer is not obliged to choose one winner for an entire subject. It can address several sub-questions and draw on different passages for each. The same cross-engine study found 78–85% of cited domains were unique to one engine, with the highest overlap only 17%. That is another reason not to bet every buyer question on one catch-all page, though the statistic is not a test of pillar pages themselves.

What I do instead is boring and specific. I read search terms the way I used to read them for negatives, except now I am collecting questions. Some ask about cost. Others ask about timelines, comparisons, failure modes, or who the service is not for. Those are not five ways to request the same generic introduction.

I group related questions, give each cluster a page that answers directly in its opening, and put the supporting detail underneath. A pricing page can explain what is included. A timeline page can say what affects the wait. A comparison page can state the tradeoff without hiding it behind a table of green check marks. The point is not to multiply pages for the sake of a page count. It is to stop making one paragraph pretend to answer everything.

Give each distinct buyer question a clear place to land. Keep the pillar if it helps people navigate; do not ask it to supply every quotable sentence in the industry.

Myth 7: “Once cited, you stay cited”

This is the one I find hardest to kill. A citation feels like a ranking you defended: win position 2, keep position 2 until someone pushes you out. But an answer can change, the passage it uses can change, and your link can vanish without a neat alert in your usual reports.

In a six-week study of 1,127 cited URLs, only 119, or 10.6%, survived all three waves. The 28-day retention rate averaged 33%; AI Overviews held at 27%, and 87% of query-platform pairs fell below 70% stability. Those figures are a warning against treating one citation screenshot as a durable asset, not a prediction that any particular page will lose its link on schedule.

The mechanism is less mysterious than the dashboard makes it look. The answer is assembled again for the query. Different wording, available passages, or a clearer competing sentence can change what it cites. Your page may still be useful and still lose that line. That is annoying, but it is not an argument for buying a more elaborate chart.

Measurement has limits, too. Search Console folds AI Overview data into aggregate totals, and GA4 has no distinct AI Overview referral parameter. I would not claim a clean AI Overview revenue number from those reports. I watch CTR changes on queries I care about, branded-search trends, and a separate log that distinguishes a cited source from a text mention. They are signals to investigate, not a receipt proving which answer caused a sale.

My weekly routine is short enough to survive an actual Monday:

1. Choose buyer questions. Pull 20 from the last 90 days of search terms. Include price, timeline, comparisons, failure cases, and who should skip you. Group related questions rather than stuffing every one onto a single pillar page.
2. Put the answer in visible copy. Rewrite the opening of each relevant page so a reader gets a plain answer first and the supporting specifics next. Do not leave a fact you want cited only in JSON-LD.
3. Check the same queries again. Log whether your page appears as a cited source, your brand appears only in the text, or neither appears. A passing mention is not a citation with the link missing.
4. Edit the passage that lost. Compare what the answer needed with what your page actually said. Tighten the first sentence, clarify an existing number where it helps, and update stale details rather than changing a date for decoration.

That routine is repetitive. I know because I used to bill hours for this kind of search work at 1am. It is also where an always-on operating model makes more sense than periodic dashboard reviews: groas handles continuous search execution while a named human strategist sets direction and guardrails, with changes logged and accountable. The strategist still has to care which buyer questions matter. The machinery should not need to be persuaded to check them again.

I had to unlearn the idea that paying, monitoring, and defending a position were one continuous plan. For citations, I write the answer, check whether it is used, and improve the passage when it is not. You do not own a citation. You rent it by the week, and the rent is paid in clearer sentences.