AI search is answering
your customers' questions.
But are you showing up?

A Denver personal injury firm brought me in because organic traffic had dropped almost a third. The managing partner opened our first meeting with one question: "Our agency says we're still ranking. So why are fewer people calling?"

He was right on both counts. Impressions were up 3 percent. Average position for the 40 searches they tracked hadn't moved. Organic clicks were down 30 percent, qualified inquiries were down 21 percent, and signed cases were down 20 percent. That's the version of this problem I keep running into. The report says you're ranking, and the phone says otherwise.

What changed is that the search engine started answering the question itself.

What actually changed

Google's AI Overviews and AI Mode, along with ChatGPT and Perplexity, don't show a list of ten links anymore. They read a handful of pages, quietly split your question into several smaller ones, and hand back one written answer with a few citations tucked into it. Most people read the answer and stop. Some click a citation. Almost nobody scrolls past it to the old blue links.

The numbers on this are not subtle. Rand Fishkin's SparkToro study of Similarweb clickstream data for January through April 2026 found that 68 percent of U.S. Google searches ended without a click, up from 60 percent in 2024. AI Overviews now show on more than 20 percent of searches, and when one appears, click-through drops by nearly 60 percent. Search Engine Land's write-up has the full breakdown.

Pew Research got to the same place a different way: in their study of 900 U.S. adults, people clicked a traditional result on 8 percent of visits that showed an AI summary, against 15 percent without one.

For a business, it means two things at once. You get fewer visitors from search.

And the visitor who does click through has already read a summary, already compared you to two competitors, and is a lot closer to picking up the phone.

One caveat worth being honest about. The same study found only a third of one percent of searches moved into Google's AI Mode.

So today, AI Overviews are the traffic problem. AI Mode, ChatGPT and Perplexity are the citation problem, and they're growing fast enough that I'd rather fix it now than explain it later. Ahrefs tracked ChatGPT referral traffic tripling between January and September 2025, growing about 14 percent a month while Google's share shrank.

Start by asking the machines about yourself

Before you change anything on the site, find out what's being said. For the Denver firm I split the traffic first and found almost the whole drop was in educational articles about fees, insurance and the claims process, down 47 percent, while attorney and service pages were flat. Those were the pages where AI summaries were showing up.

Then I built a list of 30 questions from anonymized intake themes and search data, the way a person would type them into a chat box. "How much does a Denver car accident lawyer charge?" "Who pays medical bills while an injury claim is pending?"

I ran each one through Google AI Overviews, AI Mode, ChatGPT with search and Perplexity, and repeated it on three dates. That's 360 checks.

For each one I logged the answer, the cited URLs, and whether the firm was named or cited, tracking the two separately, since an answer can mention you without ever linking to you.

I pull the starting questions from Search Console's query report and the "People also ask" boxes, then check them against Semrush or Ahrefs to see which competitor pages already hold the citations. That tells you who you're up against before you type a single prompt.

Keep the log simple. A spreadsheet with the question, the tool, a yes or no on whether you were named, what it said, and who was cited instead. That's your baseline, and it takes an afternoon.

Treat the brand as an entity, not a set of pages

Here's what I've seen make a difference. On a different job, a regional law firm I moved from a legacy CMS onto WordPress, the site had the generic LocalBusiness schema that every WordPress plugin spits out. I replaced it with custom JSON-LD that tied the firm, each attorney and each practice area to the matching Wikidata entities, added credentials and FAQ markup.

Then I went through Google Business Profile, the directories and the review sites to make sure the same facts appeared everywhere. Same name, same practice areas, same office addresses.

That consistency matters more than any single page, because the models don't trust one source. They cross-check. If your site says one thing and a directory from 2023 says another, the directory can win.

The other thing I've changed is how pages open. The firm's consultation article ranked third for "what should I bring to a car accident consultation," and the AI answer still cited a competitor, because the firm's page spent its first 500 words describing the firm.

I replaced that opening with the answer itself, then put the checklist, what happens in the meeting, who reviews the matter, and a booking button underneath.

The models lift passages, not pages. A page that takes three paragraphs to get to the point never gets quoted.

What to fix, in order

Once you have the audit, the fixes fall into a short list, and the order matters because the early ones make the later ones work.

What it did for the firm

For the Denver firm, that meant rebuilding six pages, chosen by lost traffic, relevance to signed cases and the gaps in the audit. The consultation page lost its 500-word opening and got a direct answer, a checklist and a booking button. The fees page explained the firm's actual arrangement with an attorney-reviewed example. Spanish availability went on the service, contact and attorney pages, and the stale directory profiles got reconciled.

I interviewed two attorneys so the content reflected how the firm worked, and every legal page got a named reviewer and a review date.

No special AI files, no exotic schema. Google's own guidance says the fundamentals are what apply to its AI features, and that's what I've seen too.

The other half was conversion. I simplified the mobile consultation form, tied website inquiries and tracked calls to intake outcomes.

I also added two intake questions: "How did you first hear about us?" and "What prompted you to contact us today?" That's how you catch "Perplexity mentioned you, then I Googled your name," and count it once instead of twice.

Ninety days later, here's what moved.

Measure it in money, not mentions

Three numbers tell you whether any of this is working. AI referral traffic in GA4, which you can see by filtering sources for chatgpt.com, perplexity.ai and the Google AI referrers.

Your brand mention share, meaning how many of those ten questions name you when you rerun the audit each quarter. And conversions from that traffic: calls, forms, quote requests.

The last one is the only number a client feels. I've written before about a heating and cooling company whose organic traffic rose 40 percent while the phone stayed quiet. This is the same lesson from the other side: the traffic report and the phone can disagree, and the phone is the one that's right.

It's also where the good news is. Ahrefs reported in late 2025 that AI search sent about half a percent of their visits but produced 12 percent of their signups. Buffer, cited in the same piece, saw AI-referred visitors convert at 20 percent against 7 percent for organic search.

Those visitors did their comparing before they clicked. That's the trade. Fewer visitors, better ones, and a site that had better be ready for them.

Two limits on that. Ahrefs also notes ecommerce converts worse from ChatGPT than other sectors, so if you sell products the math is different.

The businesses that gain the most here are local services and B2B, where the customer is asking a question and wants a name. That's who I'd tell to do this first.

What I wouldn't bother with

A few things get sold as AI search fixes that I'd skip. An llms.txt file on your server does nothing until a major model says it reads one, and none has.

Stuffing an FAQ block with every keyword you can think of gets you a page the models skip because it reads like a page written for models.

Trying to chase every AI tool at once burns a budget on measurement before you've fixed the facts. Get the facts right, get the openings right, then measure.

Where this is going

I think the next step is already visible. More buying decisions will start with an agent doing the research on the buyer's behalf, shortlisting three options and presenting them.

When that's the front door, nobody cares which page ranks. They care whether the model's summary of you is right, and whether you made the shortlist.

The brands that get recommended will be the ones that kept their facts clean, structured and consistent everywhere, long before the agent came looking.

That's the work I'd start on today, and it's the work most competitors still haven't started.

If you want to know how your brand shows up right now, run the ten questions this afternoon. If you'd rather have someone do it with you and tell you what to fix first, get in touch.