A ChatGPT citation is the new rankings report: a nice-looking number that can hide a silent phone. If your brand appears in answers to questions nobody asks before hiring you, you have won a mention, not an enquiry.

I judge this the way I judged every Google Ads account I ever ran: by cost per lead, not sightings. Pew tracked 900 US adults across 68,879 searches and found users clicked a traditional result on 8% of visits with an AI summary versus 15% without, and clicked a link inside the summary on only 1% of visits. Read that twice. A citation is mostly an impression. You already know what impressions are worth when the invoice arrives.

I used to send clients ranking reports back in 2015. Position three, up from five, green arrow, everyone happy, phones silent. The citation report is that same report in a new costume. The fix is not a better costume. It is to start with the question a buyer asks before money moves, then make sure the page they reach gives them a reason to get in touch.

The citation pitch sells a visible win, not a buyer

The sales call goes like this. Your Google traffic fell after AI Overviews rolled out, which it probably did. Ahrefs re-ran 300,000 keywords against Search Console data and found position one gets 58% fewer clicks when an AI Overview shows up. Reddit told shareholders search referrals fell by more than 1M per week in October, blamed on Overviews and AI Mode. So the agency offers the fix: we will get you cited in ChatGPT, Perplexity, and Google answers. Weekly report, logo wall of mentions, green arrows.

That is an easy sell because it replaces one visible loss with another visible win. Clicks down, mentions up. The problem is what those mentions measure. Most trackers count any appearance in any answer, including answers to informational prompts like “what is powder coating” that never precede a purchase. A citation with no buyer behind it is an impression you paid agency rates for. I stopped treating impressions as a win in paid search years ago. You can read why in the 7 Google Ads metrics that actually matter.

Here is the split I want in the report. An informational prompt sounds like “how does emergency garage door repair work.” A hiring prompt sounds like “who fixes a broken garage spring in Austin tonight.” Only the second one carries a wallet. When I audited citation reports earlier this year, almost all the volume sat in the first bucket. Clients were paying to win answers read by students, DIYers, and competitors.

That does not make the first question useless in every context. It makes it the wrong unit of success for a business buying enquiries. If the report puts both prompts in the same column and calls the total growth, the column is the problem. Separate buyer questions from general questions before you pay to win either one.

A source can win the answer without winning the enquiry

Google does not keep a secret second index for AI answers. It works from the index you already rank in, then runs one question through what Citepoint calls query fan-out: your search gets split into related sub-searches, each pulls pages, and the answer gets stitched together from pages that answered a sub-question cleanly. That is why a page sitting at position eight for the main term can still get cited. It did not win the query. It won a sub-question.

Because the system fetches pages and extracts short answers, the pages that get pulled share three traits: they are indexed and eligible for a snippet, they answer one thing in a quotable sentence, and they look current. No special schema or llms.txt buys your way in. I have seen dated, specific pages with a price, a service area, a step count, or a last-updated date beat a prettier brand page for the simple reason that the model can quote them without guessing. For the mechanics of making a page fetchable, read how to get cited in ChatGPT and Perplexity. Fetch first, citation second.

Here is the part citation sellers skip: on buyer prompts, your own site is often not the source. Citepoint’s October 2026 index of 314 software buyer prompts found Overviews averaged 6.4 citations, cited Reddit in 50% and YouTube in 36% of answers, and sent only 21% of citations to the recommended brands’ own sites, versus 70% for ChatGPT.

So a tidy rewrite of your blog cannot be the whole plan. The buyer may see an answer assembled from your page, someone else’s discussion, and a video. You still need a page worth quoting, but you also need the answer to point toward a business the buyer can choose. If you fix only your blog, you can lose the prompt that actually hires someone.

Start with the question asked before money moves

I used to tell clients to start from keyword volume. I was wrong. Volume tells you how many people ask. It tells you nothing about how many will pay you. Say you are spending $20k a month on ads and your organic traffic fell 30% after Overviews. The fix is not to win back all 30%. Half of it was tire-kickers the answer now satisfies directly, and no page rewrite brings them back. What the deck calls “visibility recovery,” I call chasing people who were never going to call.

Start from the five to ten prompts a buyer types right before money moves. Price, turnaround, service area, comparison, risk. “How much does dustless blasting cost per car in Dallas.” “Best garage door company for same-day spring repair.” “Inuteq vs generic cooling vests for warehouse crews.” Write each prompt on a card. Then ask two questions:

  1. Does one of your pages answer it in a sentence an AI answer can quote?
  2. Does that same page give a skeptical visitor who arrives cold a reason to enquire?

If either answer is no, that prompt belongs on your to-do list. Order the list by margin, not search volume. A broad question can produce a satisfying chart; a buyer question has to earn its place by getting someone closer to a decision.

Build the page so a model can lift one sentence and a buyer can act in 30 seconds. Put the answer in the first 80 words, in plain numbers, then give the proof. Price range, areas served, turnaround time, what happens when they call. Make the updated date visible. No essay intro, no 2,000-word glossary. Models quote what is specific and current, and buyers do much the same check with their eyes. A page that stops at “costs vary” gets skipped twice: once by the answer, once by the visitor.

The second half is the part SEO advice forgets. A quoted page that does not convert just donates a lead to whoever ranks below you. Put the call button, the form, and one reason to trust above the fold: reviews count, jobs done, response time. Do not make someone read to the bottom to discover whether you serve their area or can take the job. The page has two jobs, and being easy to quote is only the first one.

I have watched AI referrals arrive small but serious. Marie Haynes tracked 100 sites over 145M sessions and found ChatGPT sent just 0.62% of Google Organic volume but converted at 13.8% versus 3.02% for organic. Invoca logged 70M calls and found ChatGPT-referred calls became sales leads at 49%, the highest of any channel, on very low volume. Low volume, high intent. Treat that visitor like a paid click you did not have to bid on.

Ink cartoon of a marketer framing a citation report while the phone stays silent

Measure the enquiry, not the appearance

If you ran this like I ran paid search, you would ask one question: how many enquiries did it produce and what did each one cost? That kills most citation reports on contact. Google’s new Search Console report for AI answers, rolled out to all sites Aug 31 2026, shows impressions only by page, country, device and date. No queries, no clicks, no position. Clicks get mixed into main Performance totals with no AI label. Your agency can show you a chart going up while your phone stays quiet, and that report alone cannot settle the argument.

Do what paid operators do instead. Tag the pages built for hiring prompts, watch assisted enquiries in GA4, log calls that name ChatGPT or Google’s answer as the source, and tie each one to a page and a prompt. This takes more work than exporting a mention count. It also tells you what to keep building and what to stop paying for. A prompt with appearances but no enquiries is not a success story just because the line on the chart points up.

I would rather see four enquiries from three buyer prompts than 400 mentions across 90 informational answers. That is the standard I use in AEO agencies ranked by the number that matters: enquiries, not appearances. groas runs SEO and AI search that way, with pages shipped daily and every action logged with reasoning, for one flat fee. If the report cannot connect a mention to a lead, it is a hobby report.

Write off the visits that were never going to call

Some of the traffic you lost is gone for good, and trying to win it back burns money. When the question was “what is powder coating” or “how long does a garage spring last,” the AI answer settles it on the page. That visitor was never going to call. Publishers learned this first. Referral traffic from Google fell 43% year over year for large publishers and 59.8% for small publishers from June 2025 to 2026, alongside Reddit losing more than 1M search referrals per week in October. I would not pay to recover those visits for a service business. They were readers, not buyers.

Fight where the answer still needs a provider: price, turnaround, service area, license, warranty. Those answers cannot end without a name and a number, and the model has to cite someone specific and current to finish the sentence. That is the traffic you can win back with dated, local, numbers-first pages and third-party proof on Reddit and YouTube. This is not a promise to restore every lost session. It is a choice to spend effort where a session might become a job.

If the prompt can be satisfied without hiring anyone, write it off. If it ends with someone getting hired, that page gets the work. There are exceptions to who should take this advice. If you sell ads against informational pageviews, you live on volume and my approach starves you. Same for affiliate sites where the click is the product. And if you are booked out for six months, do not pay anyone to get chosen. Everyone else selling jobs, installs, and retainers should stop counting mentions and start counting enquiries.

Make anyone selling AI visibility agree to this

An agency that sells you mentions will always find more mentions to sell. There is an endless supply of informational prompts with zero buyers attached, and a tracker will count every one as progress. A PPC operator would never accept that deal. You would ask what each lead cost, which keyword produced it, and what to cut next. Hold AI search to the same standard or stop paying for it.

If you hire anyone for this, make them agree to three things in writing:

  • Name the hiring prompts first, in your service area, in order of margin, not volume.
  • Ship one quotable, convertible page per prompt, with price, area, turnaround, and proof above the fold.
  • Report enquiries by prompt and page, with calls logged, not mention share and not impressions.

Anything short of that is the 2015 rankings report again. The businesses that keep buying appearances will have beautiful charts and quiet phones. The ones that get chosen on the prompts that hire someone will take the calls they used to miss.