Stop Hunting for the Right Prompts. Our Most-Fetched ChatGPT Page Was a Changelog

Our plain Google Ads changelog drew 569 verified ChatGPT-User live-answer fetches in 30 days. It was not built around a prompt. It was a dated page about what changed and what to check, and ChatGPT could fetch it when it needed it. That is a better starting point for citation work than hunting for the right prompts.

I spent years watching advertisers chase the same kind of shortcut in Google Ads. A new match type trick. A new bid hack. Now a prompt list that supposedly triggers citations in ChatGPT, Perplexity or Google AI Overviews. The useful question is less glamorous: when an assistant looks for an answer, can it retrieve a page that gives one? Prompt research does not solve that problem.

The prompt-list advice sounds diligent. That is why it sells.

The current playbook is easy to recognise. You sort buyer prompts into definition, comparison and selection buckets, pull 30 to 60 from sales calls and autocomplete, then track which ones trigger citations. Build a sheet. Watch the mention score climb. I used to tell clients a version of this for keywords. I was wrong about how durable that work is.

It feels like keyword research with a new label, and it gives an agency something to bill for every month. The problem is mechanical, not motivational. Prompts are infinite. Fetchable pages are scarce. A buyer can phrase the same Google Ads question 200 ways. Writing a thin page for every variation does not help if the assistant cannot retrieve any of them, or if none gives it a complete answer. Chase the phrasing and you chase noise. Build the page worth retrieving first.

The distinction matters because a prompt is only the start of the trip. The assistant still has to find a candidate page, fetch it and use what is on it. A spreadsheet can tell you how someone might ask a question. It cannot make your answer available at the other end.

Same buyer question, different engine, different sources

An October 2026 test of 31 agency buyer prompts across five engines found ChatGPT with web search cited vendors’ own sites for 119 of 210 citations, while Perplexity cited listicles for 395 of 561. Those figures describe that test, not every category or future search. But the contrast is enough to puncture the idea of one universal prompt playbook. The buyer intent stayed similar; the sources the engines used did not.

The retrieval step comes before the citation

ChatGPT Search and Perplexity take different routes to web content. That is why I would check whether a page can be found and fetched before grading it against a list of phrasings. If an assistant cannot retrieve your page, it cannot cite it. If it can retrieve the page but the answer is buried or out of date, matching a prompt still will not fix the page.

This does not mean buyer questions are useless. They tell you what the page must answer. They are a poor substitute for checking whether the page is available to the systems you want to reach. Practical takeaway: use questions to shape the answer, not as proof that an assistant can see it.

What our edge logs showed, and what they did not

The number that stopped me scrolling was 569 verified ChatGPT-User live-answer fetches for one Google Ads changelog in 30 days. The second page was another dated what changed explainer, with 402. Neither was written to match a catalogued prompt. Both gave an assistant a current, specific page to pull when it needed information about a change.

Raw bot counts would have made that story less convincing. A user-agent string can be spoofed, so we filtered for requests whose source IP matched OpenAI’s published ranges and whose path indicated a ChatGPT-User live-answer fetch. Our separate breakdown found 76.5% of claimed AI bot hits failed verification. Counting every request that says GPTBot would be easier. It would also tell us much less.

Verified fetches are a signal, not a citation count

Here is the useful comparison from the 30-day edge-log analysis:

A live-answer fetch is not proof that the finished answer cited the page. It is a verified request made while ChatGPT-User was retrieving content, not a screenshot of what the reader saw. This is also one site, one 30-day window and a truncated path list. The logs do not show that changelogs win in every niche, that these pages generated enquiries, or that Perplexity and AI Overviews behave the same way. They show something narrower: two useful maintenance pages drew 971 verified live-answer fetches between them without being built around a prompt list.

That is enough to change the order of work. Before paying for another mention report, I want to know which pages an assistant is actually requesting and whether those pages answer the question. Practical takeaway: verify the request, then copy the page’s usefulness—not its topic.

Do the work a prompt sheet cannot do

Forget the prompt sheet for a month and fix the fetch path. Citation work has four jobs, in order:

1. Let the relevant bots find and retrieve the page.
2. Publish an answer worth retrieving.
3. Keep that answer accurate as things change.
4. Measure whether the attention leads to business.

That is shorter than any prompt catalog. It also gives you something to inspect when results disappoint, rather than another column to add to the sheet.

First, check whether the page is available

Start with the plumbing because it can disqualify a page silently. PerplexityBot and Perplexity-User have different retrieval roles. Check the access you allow each rather than treating a user-agent name as proof of what happened. For Google AI features, search eligibility and snippet availability matter; extra schema is not the gate. Restrictions such as noindex or nosnippet deserve a deliberate check before anyone starts rewriting copy for AI Overviews.

On a site, I would inspect three things before writing another page: whether bots can get usable HTML without working through a maze of JavaScript, whether the page is present in the relevant search indexes, and whether snippet settings allow the page to be shown. Then I would check the logs for actual retrieval. Our fetch checklist for ChatGPT and Perplexity goes deeper, but the order is the point. Do not optimise a page you have not made available.

Next, publish the dated answer people keep needing

Every industry has a question that comes back. In paid search, it is what changed and what do I check first? In home services, it may be price and lead time. In SaaS, it may be what broke after a platform update. The asset does not have to be a changelog if nothing changes. It has to answer the recurring question without forcing a reader—or an assistant—to assemble the answer from ten thin pages.

For a page about changes, put the date where it is easy to find. List what changed, explain why it matters, give the steps to check it, and say what the page does not cover yet. That structure is not a trick for one prompt. It is useful to the person who lands there, too. Our changelog had the advantage of specificity: an assistant looking for an update could fetch one page with dated entries and practical checks. A definition post written because a prompt tool found a phrasing has to earn that usefulness on its own.

Keep it current, but do real updates

Freshness gives a changing page a reason to be fetched again. One analysis of AI citations found a skew toward more recently published or updated content compared with Google organic results. Its figures describe the material it analysed, not a guaranteed lifespan for your page. They are a reason to maintain an updates page, not to change the date at the top and call it work.

A real update adds a new entry, revises a step or removes a workaround that no longer applies. If you cannot sustain that work, do not promise a monthly changelog. Publish a stable answer to a stable question instead. Freshness you can maintain beats a schedule you abandon.

Finally, measure enquiries rather than mentions

Prompt tools sell a mention chart that can climb while the phone stays quiet. I have seen that movie in PPC: click volume rises, qualified pipeline falls, and someone calls the report a success. For citation work, I want the chain that can eventually pay: verified live-answer fetches by page, sessions or enquiries you can attribute, then calls, form fills and closed revenue where the data connects them.

Our logs establish fetches. They do not establish that those fetches produced customers, and I would not put a revenue number beside them until we could connect it. A citation that never produces an enquiry may still be useful, but a mention score alone cannot tell you whether it was. Keep the fetch log and the CRM source field. Make the prompt report earn its place.

Who should ignore the changelog idea

A changelog is not the answer to every recurring question. If your Perplexity visibility depends on third-party listicles, the agency-prompt test’s listicle-heavy results give you a reason to pursue inclusion there. That does not make a prompt sheet your primary asset. It means the page worth influencing may belong to someone else.

If your product barely changes and buyers keep asking the same stable questions, build a tight FAQ rather than a freshness treadmill you will abandon. And if you cannot make substantive updates, do not put a new date on stale instructions. Buyers notice when the steps no longer work, even if a bot does not. Match the asset to the question, then maintain the asset you chose.

Publish the page worth fetching, or keep buying prompt reports

The lesson from our logs is not everyone needs a Google Ads changelog. It is that two dated maintenance pages drew 971 verified ChatGPT-User live-answer fetches in a month without targeting a prompt list. Fetches are not citations or revenue. They are evidence that the pages made it through a step prompt research cannot do for you.

Pick the recurring question in your trade. Make its answer available, specific and worth keeping current. Check who fetches it, then check whether the attention turns into enquiries. groas runs that kind of execution loop for clients because execution is the product; a spreadsheet of prompts is not. Ignore the retrieval problem and keep arguing about phrasing. Your competitor can spend that time publishing the page an assistant can actually use.