September 22, 2026
I Tested Whether AI Cites Boring SEO. It Does. Mostly.
Seven real client keywords, four AI engines, one afternoon. The engine that reads the open web cited the client every time. The engines that read Google's own surfaces cited the map pack and the Better Business Bureau. The work that earns a citation turns out to be the work you were already supposed to be doing.
AI citations are a byproduct of ranking. That is the theory. I checked.
The theory going around the practitioner side of SEO right now is that AI visibility is not a discipline. It is a side effect. Fix the technical debt, build the pages people transact on, earn links every month, and the citations show up on their own. The people selling GEO retainers disagree, which is the most predictable sentence I have ever typed.
I like theories. I like receipts more. So I pulled seven keywords where a client of mine holds a top-five organic spot in Semrush and asked Google AI Overviews, Google AI Mode, Perplexity, and ChatGPT search the same seven questions. A dentist, a car wrap shop, and a kitchen remodeler. Real businesses, real rankings, no prompt engineering. (The prompt was the keyword. Crazy to test the thing the way a customer would use it, I know.)
Twenty-three answers came back. Google showed an AI Overview for two of the seven queries. The other five got ten blue links and nothing else, which felt nostalgic, like finding a phone book in a drawer.
What four engines did with the same seven questions
Perplexity cited the client's own site in seven of seven answers and named the business in six. It pulled 18 to 32 sources per answer, and the client's page was in the pile every single time. If you rank, Perplexity finds you. That is the whole rule.
Google AI Mode cited a client site once. Once. For the other answers it listed twenty or more sources per question, and twenty of them pointed at google.com itself. Business Profiles. Maps listings. AI Mode was not reading websites. It was reading the map pack out loud, slowly, like a relative reading a menu.
ChatGPT search cited a client site zero times out of seven. For "bathroom remodel lincoln ne" it returned thirteen sources: the Better Business Bureau five times, Houzz three times, then Expertise, BuildZoom, Yellow Pages, and Angi. Thirteen citations about remodelers and not one remodeler. It did name the wrap shop for "car wraps dallas", complete with a 5.0 rating it lifted from a directory, not from the shop's site.
AI Overviews showed up twice and cited a client once, for "tuxedo wrap truck". The client's page ranks fourth for that term, and the page answers the question a human would actually ask in its first sentence: what does it cost. Tuxedo wraps start at $1,800. That one page got cited by the Overview, by AI Mode, and by Perplexity. Three engines, one boring service page with a price on it. No schema trick. No llms.txt. A price.
Ranking is the ticket. The directories are the seat.
Each engine reads a different library. Perplexity reads the open web, so organic rank is the whole game there. Google's AI reads Google, so the Business Profile, the reviews, and the map pack are the whole game there. ChatGPT reads the corroborators: the BBB, Houzz, Angi, Yelp, whoever has been vouching for you since before it was cool. Three libraries, three cards. You need all three.
So the boring list is longer than the boring list you had. Technical health, transactional pages, monthly links, and then the part everyone skips: a complete Business Profile and the directory listings you stopped updating in 2021. (The BBB listing is still there. It has been waiting. It has a lot of time on its hands.) None of that is new. That is the point. The engines did not reward a new tactic. They rewarded finished work.
One more thing the test caught. Semrush's database has the remodeler at number one for "kitchen remodel utah county". A live desktop pull from outside Utah did not show it in the first nine results. That is not a Semrush problem. Local queries move with the searcher's location, and a rank tracker is a map, not the territory. Run the live query from the city before you put the number in a deck.
The playbook that earned those citations, in order
One. Kill the technical debt, and check that it stays dead. Two real ones from my own client fleet this summer: a page-speed plugin that stripped the structured data off every page each time it optimized the site, and a caching layer that kept serving a stale bundle of JavaScript after a deploy. Neither showed up in Search Console. Both showed up in what the engines could see. The site was fine. The version of the site the machines got was not.
Two. Build the transactional page for the term that is sitting right there. The remodeler ranks 13th for "bathroom renovation lincoln", 1,900 searches a month at a keyword difficulty of 12, with no page built for it. It ranks 10th for "utah kitchen remodel", 320 searches a month, difficulty 4. Difficulty four. That is not SEO. That is remembering to make the page.
Three. Earn links every month, not every quarter when someone remembers. The remodeler's profile today is 5,154 backlinks from 484 referring domains and an Authority Score of 14. Fine for a remodeler. Not enough to hold the ticket in every city it operates in.
Four. Finish the Business Profile and the directories, because that is what two of the four engines actually read. Then measure the citations as an outcome, not a workstream. A survey of 131 SEO experts published this month put relevance (57.1 percent), backlinks (54.8 percent), and content quality (47.6 percent) at the top of the ranking factors. Semrush's own study of 481 marketers found only 22 percent have SEO and AI search fully integrated, and the integrated teams were far more likely to report traffic or leads from AI platforms. The people treating it as one job are winning. The people running two decks for the same work are paying twice.
One toolkit for the loop, because the loop is the strategy
The loop above is four tools I already run every month, and they live in one place. In the Semrush SEO Toolkit, Site Audit is how I caught the stripped structured data (it flags the pages where the markup that should be there is not). Keyword Magic Tool is where "bathroom renovation lincoln" at difficulty 12 was hiding in plain sight. The Link Building Tool turns the monthly outreach into a queue instead of a mood. Organic Research shows where the competitor with fewer keywords is out-earning the client, so the budget goes to signals that already work instead of the newest acronym.
Then the outcome check. The AI Visibility features in Semrush One track whether the rankings are turning into mentions and citations across ChatGPT, Perplexity, Gemini, and AI Mode, which is exactly the test I ran by hand above, except it runs every week and does not need me to remember.
The engines did not reward the clever work. They rewarded the finished work.
If you want the version of this test aimed at what ChatGPT does with a brand name, it is in the shortlist piece. If your problem is upstream of all this, pages Google crawled and never indexed, start with the indexing field guide.
Seven keywords is not a study. It is an afternoon. But it is an afternoon with the same result every engine kept handing back: rank on the open web, be complete on Google's own surfaces, be corroborated by the directories, and the citations arrive without a retainer. Boring is not a strategy. Finished is.