The AI Search Visibility System: Getting Cited by ChatGPT Is a Marketing Engineering Job

Billions of people now read an answer before they see a link

Your buyer has a question about what you sell. More and more often, the first thing they read is not your page; it is a paragraph an AI wrote, with a few links under it.

The scale is not small. Google said in June that AI Overviews has more than 2.5 billion monthly users and AI Mode more than one billion, and in September OpenAI put ChatGPT at 1.2 billion weekly users, The Verge reported.

Whether those answers name you is now a marketing question with a weekly number attached. It is also one of the six systems a marketing engineer builds.

Greg Isenberg puts the question in his episode on the role as whether these AI systems can make sense of your company at all, with agents pulling the data and improving the site. We agree with the frame, and this article is the build.

Here is the short version. Fix access, ask the same buyer questions every week, count how often you are named, rewrite one page, and measure again.

What our own runs taught us, in one paragraph

We have done this for our own market. On September 22, 2026, two runs read 409 AI answers to unbranded buyer questions and logged 4,243 citations.

Company pages were 65% of what got cited, "best of" posts 26%, Reddit 1.0% and YouTube zero. The full breakdown, by engine and by source type, is in What AI Answer Engines Actually Cite.

The lesson for this article is not the mix. It is that one run is a photograph, and the engines keep moving.

That same test found ChatGPT and Perplexity mostly cited different websites. A plan built on one engine's favorite source ages fast, and the only defense is to keep measuring.

Why a single check tells you almost nothing

Ask ChatGPT the same question twice and you will often get two different lists. SparkToro measured how different.

In research published in January 2026, Rand Fishkin and Patrick O'Donnell had volunteers run 12 recommendation prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. They report less than a 1 in 100 chance that ChatGPT or Google's AI gives the same list of brands twice, and about 1 in 1,000 for the same list in the same order.

Their conclusion changed during the work. Rank in an AI answer is close to meaningless, they wrote, but how often a brand appears across many runs of similar prompts held up as a reasonable metric.

That is what this whole system rests on: rate, not rank. Count how often you are named across many runs, and never report a position.

One screenshot of a missing answer proves nothing. Neither does one screenshot of a win.

Fishkin also tells buyers of tracking tools to ask for the method behind the numbers. We would add one question: how many times was each prompt run?

Step zero: make sure the engines can read you

No tracking matters if the crawlers are blocked. Check this first; it takes about ten minutes.

Each engine runs its own robots. OpenAI's crawler documentation says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, and it treats GPTBot, the training crawler, as a separate choice.

Anthropic's crawler page says blocking Claude-SearchBot may reduce your visibility in Claude's search results. Perplexity's crawler page recommends allowing PerplexityBot if you want to appear in its results.

Google adds a switch of its own. As of August 31, 2026, every site has a Search Console control for AI Overviews and AI Mode, set to include by default, and Google says sites that opt out get no traffic or impressions from those features.

Then check the layer under all of it. A firewall or bot-protection rule can block a crawler that robots.txt allows, which is why Perplexity's page walks through firewall settings.

Training and search are separate choices. You can say no to the training crawlers and yes to the search crawlers; just do it on purpose, line by line.

The Mention-Rate Loop: ask, log, count, fix

The loop runs every week. The figure is a step plan you can copy.

Ask. Keep a fixed set of 25 to 50 unbranded questions a buyer asks before they call, and run each one several times on each engine you care about. Fixed matters: if the questions change every week, the trend line means nothing.

The questions come from your own buyers, not from a keyword tool. Sales calls, support tickets and Search Console queries are the source, which is the job of a customer truth system.

Log. Save every answer and every cited link into one table, with the engine, the date and the run number. The figure in the next section lists the columns.

Count. For each question and engine, the mention rate is the share of runs that name you or cite your page. Track it next to the share of citations going to each source type: company pages, comparisons, directories and forums.

Fix. Pick the one question with the most buyer value and the lowest rate, and rewrite or build the page that should answer it. Then let next week's run say whether it worked.

One page a week is slow on purpose. It is the only pace at which you can tell which change moved the number.

Here is the math with made-up numbers, so the size of the job is clear. Thirty questions, four engines, five runs each is 600 answers a week, which is a job for an agent, not an intern.

Say you are named in 42 of the 600. Your overall mention rate is 7%, and that is the number for the Monday memo.

The useful number sits one level down. One question, the one your best buyers ask, might name you in 0 of its 20 runs while a competitor shows up in 15.

That question is this week's page. If the rewrite works, you might see 4 of 20 a few weeks later, and you will know which page did it because it was the only one that changed.

Small counts swing, so do not celebrate one week. Read the rate over four weeks before you call a rewrite a win or a loss.

What to log, and the free reports that help

You do not have to collect everything by hand. Two of the big platforms now report AI citations to site owners for free.

Google's generative AI performance report in Search Console shows impressions for your pages in AI Overviews and AI Mode, with pages, countries and dates. Google announced it in June 2026 and says the same data also stays inside the overall performance report.

Microsoft's AI Performance report in Bing Webmaster Tools, launched in February 2026, shows citations across Copilot and Bing's AI summaries, which of your pages were cited, and sample grounding queries: the phrases its AI used to find your content.

Both are counts, not verdicts. Microsoft's own help page says the report does not measure rankings, authority or importance.

For ChatGPT, Claude and Perplexity we have not found a comparable report for site owners. So you run the questions yourself or pay a tracking vendor.

Vendors exist for exactly this. Profound, the AI search company that named the marketing engineer role, sells this kind of tracking, and so do others; judge any of them by the method they publish.

What a page that gets lifted looks like

When the loop says a page is losing, the fix is usually plain. The answer sits too low, the page says what everyone else says, or it has no numbers of its own.

Google's guide to its generative AI features explains why. Its AI answers are grounded in pages retrieved from the normal Search index, and it uses query fan-out, where the model runs related sub-questions to gather more results.

So a page that answers the main question and the obvious next three has more ways in. Google also warns that making a separate page for every fan-out variant, mainly to manipulate results, breaks its scaled content abuse policy.

The same guide asks for non-commodity content: a first-hand view or real experience, not a summary of what is already out there. A page with your own dated numbers is the clearest version of that.

In the one controlled study we know of, adding citations, quotations and statistics made pages more visible in AI answers. Keyword stuffing did not help.

Google's guide is just as useful for what it says to skip. For Google Search, it says you do not need llms.txt files, special schema, chunked content, or text rewritten just for AI.

A short build list, then: the answer in the first paragraph, your own numbers with dates, a source for every claim, a named author, and clear headings. Our E-E-A-T checklist covers the trust half, and hub-and-spoke linking covers how the page connects to the rest of the site.

"Google says AEO is not a real thing."

For Google Search, that is roughly what Google says, and we agree with it. Its guide calls many AEO and GEO hacks ineffective and says the foundations of SEO still apply.

Nothing in this system is a hack. Access, measurement and better pages are ordinary marketing work, done on a weekly clock.

Two things still make it its own job. ChatGPT, Claude and Perplexity run their own crawlers, so a site can be open to Google and closed to them.

And the answers change from run to run, so you cannot see where you stand by looking once. You have to count.

Where your buyers ask in the first place also differs by topic. Our Ask-Ratio analysis found one topic asked of AI assistants 3.4 times as often as it was searched on Google, and another searched eleven times as often as it was asked.

Who runs it, and how it fits the week

This is a marketing engineering system because most of it is plumbing. An agent runs the questions, saves the answers and fills the table; a person picks the page and approves the rewrite.

Give the agent a written brief: which questions, which engines, how many runs, where the table lives, and what it must never do, such as publish anything. The mention rates then become one line in the Monday growth cockpit memo.

Expect slow numbers. A rewritten page has to be crawled again before any engine can cite it, and OpenAI's crawler page notes that even a robots.txt change can take about 24 hours to reach its search systems.

Google's AI Mode is also still changing fast. Google says AI Mode queries more than doubled every quarter in its first year, so last quarter's baseline is already old.

If you want the system built inside your company, that is what our marketing engineering engagements do, one system at a time. If the page work itself is the gap, our SEO service handles it.

What to read next, and why

The Page-2 Problem [case study]. Our own pages sat on page two of Google, seen and not clicked. AI engines retrieve from that same index, so read this before you blame the AI.

Programmatic AI-for-Industry Pages Without a Google Penalty [warning]. The line between a useful page per question and scaled content abuse. Read it before you build fifty pages for fifty prompts.

The Marketing Engineering Audit [tool]. Twenty questions on where your growth systems stand, AI search included.

How We Run a Marketing Agency as Marketing Engineers [proof]. The tracking runs, the checks and the content engine behind this library.

So the simple takeaway is this: you cannot rank in an AI answer, but you can count how often it names you, and you can move that count one page at a time.