Somewhere in your Search Console query report sits a string like "what's the best way to migrate a 50-person team off spreadsheets without losing three years of formulas." Nobody "googles" like that. Somebody prompts like that. ChatGPT search prompts in Google Search Console became one of the stranger SEO artifacts of the past year, and understanding where they come from turns a curiosity into two genuinely useful signals, plus one cautionary tale about measuring a system mid-rewire.
This is the explainer: origins, the lawsuit that changed the plumbing, and what the strings do and don't mean. The companion piece covers how to use them without overfitting; this one covers what you're actually looking at.
Where Prompt-Shaped Queries Come From
Two currents feed the phenomenon, and your reports blend them without labels.
People learned to search like they prompt. After a few years of chatbots rewarding full sentences with context and constraints, the habit followed users back to the search box. Query-length drift shows up in ads data and practitioner reports alike: more words, more qualifiers, whole questions where fragments used to live. This current is organic, durable, and growing, and it's the one that makes the strings valuable, because it's your actual market speaking in its actual voice.
Machines passed through Google's ecosystem. The stranger current. AI-mediated journeys leave residue: assistants that search on a user's behalf, tools that verify answers against Google, and, most notably, the period when ChatGPT's own retrieval reached Google's results through intermediaries. When a machine's synthesized sub-query lands on Google's surfaces, it can look, in reporting, like a very odd human. Some of the most unnatural strings in 2025-era GSC reports were plausibly never typed by anyone with hands.
You can't cleanly separate the currents in your own data, and for practical purposes you rarely need to. But the second current has a story worth knowing, because it just changed shape in court.
The Scraping Era, The Lawsuit, And The Rewire
Here's the arc, per court filings and practitioner telemetry. ChatGPT's retrieval stack was never just one index: it has drawn on Bing, its own growing index, and, historically, Google results obtained through third-party SERP API services. Telemetry work identified the internal retrieval engines by name, including OpenAI's own index alongside connectors matching commercial scraping providers. In December 2025, Google filed suit over the scraping route. Since then, measurements show the stack leaning on Bing and OpenAI's own index, with the Google-scraping path shut down.
Two consequences matter for SEO. First, the machine-residue current into GSC weakened: fewer of ChatGPT's hidden sub-queries touch Google's surfaces, so the strangest strings thin out over time, and what remains skews back toward the human current. Second, and bigger: Bing quietly became load-bearing again. Early measurements had found the overwhelming majority of ChatGPT search citations matching Bing's top results; that overlap has fallen substantially as OpenAI's own index took over, but Bing remains wired into the stack serving both ChatGPT and Copilot. The practical takeaway fits on a sticky note: if your pages don't exist in Bing, part of ChatGPT can't find you, and almost nobody checks.
There's also a quiet lesson about measurement in that arc. An entire genre of "ChatGPT keyword data" was, for a while, partially an artifact of one company scraping another, and it changed character the day the plumbing changed. Instruments in this field measure the pipes as much as the water, which is exactly why single-instrument strategies mislead.
200+ SaaS teams already track their AI citations.
They know exactly when ChatGPT mentions their brand, and when it stops. Do you?
The Timeline, For The Record
Because the mechanism keeps shifting under the reporting, here's the dated arc in one place.
Early 2025: the Bing era. Measurement of the then-new ChatGPT search found 87% of its citations matching Bing's top organic results. Optimizing for ChatGPT effectively meant ranking in Bing, and GSC's prompt-strings were mostly the human current.
Through 2025: the plural era. Telemetry identified a retrieval stack drawing on OpenAI's own index plus connectors consistent with commercial SERP-scraping services reaching Google's results. This is the window when the strangest machine-shaped strings peaked in Search Console reports.
December 19, 2025: the lawsuit. Google filed over the scraping route. The intermediary path wound down.
2026: the own-index era. Estimates put the ChatGPT-citation-to-Bing overlap near 27% as OpenAI's index took over the bulk of retrieval, with Bing still wired in as a secondary. Then August's GPT-5.6 update rewired selection again, scoping retrieval toward trusted domains, which changed who gets cited without changing where the fetching happens.
Read the arc and one meta-lesson stands out: every stage rewarded the same underlying assets, indexed pages, extractable answers, and a verifiable brand, while punishing whoever had optimized for the previous stage's plumbing. The stack will rewire again. The assets won't change.
How often does ChatGPT mention your brand?
Most founders have no idea. The answer might surprise you.
Telling The Currents Apart In Your Own Report
You can't separate human from machine strings perfectly, but the forensics are better than nothing, and twenty minutes with your export teaches you your own mix.
Machine-shaped tells: qualifier stacking beyond human patience ("for teams of 40-60 in regulated industries with existing Salesforce instances"), template-like structural repetition across otherwise different queries, perfectly grammatical forty-word constructions with zero typos, and strings that read like a system prompt's paraphrase of a question rather than the question. Human-prompt tells run opposite: first person ("should I..."), typos and abandoned syntax, emotional qualifiers ("that doesn't suck", "before my boss asks"), and the conversational shorthand people use when they think nobody's grading them.
The punchline, though, is how little the distinction ultimately changes your actions. A machine-shaped string still encodes a real user's intent one synthesis step upstream, and a human prompt is the intent raw. Both feed the same clustering, both reveal the same constraint vocabulary, and both reward the same answer-shaped pages. The forensics are worth doing once, to calibrate how AI-mediated your category already is, and that calibration is genuinely useful: a report thick with machine-shaped strings means the agent-driven journey has already arrived in your market, and the agent-readability work just moved up your priority list.
What The Strings Actually Tell You
Strip the mystique and the prompt-shaped queries carry two honest signals and two temptations to resist.
Signal one: phrasing intelligence. Full-sentence queries expose the constraint vocabulary buyers actually use: team sizes, stacks, budget shapes, deal-breakers. That vocabulary belongs in your comparison pages and FAQs, in your words but their frame. It's focus-group data at zero cost.
Signal two: presence evidence. Your pages surfacing for prompt-shaped searches means you're already in the retrieval paths of AI-mediated journeys on Google's side. That's a floor, not a ceiling, and it says your content is legible enough to appear when the question gets specific, which is where deals live anyway.
Temptation one: treating strings as keywords. A forty-word phrasing appeared once and may never recur. The unit of action is the intent cluster, and the overfitting guardrails exist because this temptation produces four hundred thin pages before anyone notices.
Temptation two: treating GSC as an AI visibility report. The strings prove prompt-shaped demand exists on Google's surfaces. They say nothing about whether ChatGPT, Perplexity, or Claude cite you inside their answers, which is where the shortlists form. That question has no Google-side instrument at all; it takes per-engine citation tracking on a fixed query set, run weekly because answer sources churn hard between checks.
What This Means For Your Free-Tier Reality
One more mechanical detail connects the strings to your content decisions, and it's the one with the highest payoff per word changed. Most ChatGPT usage runs on the free tier, and measurement of free-tier behavior found the majority of its answers assembled with zero page-opens: the model works from your title, URL, and roughly the first 200 characters of body text as stored in its index. Now put that next to the prompt-shaped queries in your GSC report. The people asking forty-word questions are being answered, on the other side of the fence, from two-sentence excerpts.
That's the whole case for answer-first openings in one image. The market asks long; the machine quotes short; the page that wins is the one whose shortest excerpt answers the longest question. Audit your top pages' opening lines against the constraint vocabulary your prompt-shaped strings contain, and close the gap sentence by sentence. It's the rare optimization where the evidence for the demand and the surface being quoted are both sitting in instruments you already have open.
The Short To-Do List
Four moves convert the explainer into action. Verify your key pages in Bing by name and by topic, this week, because the post-lawsuit stack made it plumbing again. Cluster your prompt-shaped strings monthly and strengthen one deep page per recurring intent, speaking the constraint vocabulary the cluster surfaced. Keep your opening paragraphs answer-first, since prompt-shaped searchers, human or machine, get served from exactly that text. And stand up the per-engine measurement that GSC will never provide, so the next time the plumbing changes, and the lawsuit proved it changes overnight, you're reading your own data instead of archaeology.
The strings in your reports are a small window into a big migration: the search box learning to speak chatbot, and the chatbots renegotiating whose index they read. Watch the window, act on the clusters, and keep your instruments plural, because in this era the pipes get sued, and the water finds new ones.
15 hours a week manually. Or 15 minutes with RankControl.
Track citations, monitor competitors, and fix content gaps across every AI search engine. Automatically.
