Most citation advice is a guess with good formatting. Before you rewrite another page to satisfy someone’s list of AI search rules, spend 30 days finding out which engines cite your site, for which prompts, and which pages they pull.

I wrote 7 Myths About Getting Cited by ChatGPT and AI Overviews to clear away false beliefs. But knowing what is false does not tell you whether AI Overviews, ChatGPT or Perplexity cite you today. They pick sources differently, and what they pick changes. This is a second look at the question, this time with a sheet you can run yourself.

The test has one question: for a fixed set of buyer prompts, where do you get cited on each engine, and which of your URLs appear? Record a visible citation or link separately from a brand mention without a link. Record a relevant bot fetch separately from both. A fetch tells you a page was reached; it does not tell you that the page appeared in the answer. Do not roll those three observations into one visibility score.

My expectation is an uneven sheet: cited on Perplexity for a comparison prompt, absent from AI Overviews for the same question, perhaps fetched but never named on ChatGPT. Different retrieval and citation patterns can produce that split. Keeping the engines separate is how you find out where to spend the next hour.

First, set the budget and the boundary

You cannot buy an organic AI Overview citation. Google’s Ads Help says ads in AI Overviews and AI Mode are labeled Sponsored and visually distinguished, and that paid advertising does not influence organic AI responses. An ad beside an answer and a citation inside it are separate things. One you bid for. The other is what this test measures.

The same Google document says those ads use the Search auction and CPC and conversion bidding, with no extra charge, and that you cannot target only AI Overviews or AI Mode. Budget zero extra for citation placement. Budget time for measurement: about three to four hours to build the prompts and sheet, then roughly an hour a week to run prompts and record what happened, plus whatever you already pay for server log access. Free tiers can get you through week one if cash is tight; clean sessions and repeatable settings matter more than buying another tool.

I used to tell clients to fix tracking before touching bids. Same principle here. If you cannot tell a citation from a mention or a training crawl from a live fetch, you are not ready to spend a month rewriting pages.

Setup: write prompts buyers might actually use

Pull 30 to 50 prompts from sales calls, support tickets and queries already sitting in Search Console. Keep the customer’s wording, including awkward phrasing and typos. If you sell garage doors in Dallas, include both “garage door repair cost Dallas” and “my garage door broke at 11pm who actually answers.” Thirty prompts you repeat beat 200 you run once.

Build the sheet before you open an engine. Give each prompt an ID so you can compare the same question across runs. For every run, record:

  • Date, engine, prompt ID and exact prompt text.
  • Whether your site received a visible citation or link, and the URL cited.
  • Whether your brand appeared without a link.
  • The URLs of the two competitors cited most often.
  • Whether your logs show a relevant fetch near the run time, marked as a separate observation.

Do not mark an answer “cited” because it sounds like your page. You need a visible citation or link you can record. If the answer names your brand but links elsewhere, put that in the mention column. The distinction will matter when you decide whether to work on retrieval, the page itself or neither.

Tag each prompt by intent: definitional, comparison, how-to, local or price. Then choose ten prompts for a phrasing check. Run each as a short keyword query, such as “best CRM for plumbers,” and as a longer conversational question, such as “what CRM should a 5-person plumbing company use if nobody will do data entry.” Keep both versions in the sheet as separate rows. You are testing whether phrasing changes what your buyers see, not trying to discover a universal prompt incantation.

Use a clean profile, log out where you can, and keep the location setting consistent. Write down any setting you cannot hold constant. The controls will not make three different engines identical. They will stop you from mistaking a changed session for a changed citation pattern.

Before week one: give each bot its own column

Open your server logs and filter by the bot’s job, not the word AI. OpenAI lists separate identities: OAI-SearchBot for surfacing sites in ChatGPT search, GPTBot for training crawls, and ChatGPT-User for fetching a page when a user asks for it. Perplexity distinguishes PerplexityBot, which surfaces and links sites in results, from Perplexity-User, which fetches after a user question.

Keep discovery crawls, training crawls and fetches near your test runs apart. Otherwise you can end up celebrating a training crawl as if it were a citation. Check the bot identity rather than trusting a user-agent string on its own. Note the time and URL of any relevant fetch, but do not claim you have tied that request to a particular answer unless your evidence supports it. The answer on screen remains the citation record.

Leave llms.txt out of the intervention. In an analysis of 137,000 domains, 97% of those files received zero requests in May 2026, and live-answer bots accounted for 1.1% of hits. Your first question is simpler: did a relevant bot reach the page you hoped would answer the prompt? If not, polishing a sentence on that page is unlikely to solve the access problem.

Week 1: run the baseline twice

Run the full set on each engine, using the same setup for every prompt. Three days later, run it again. Record the URLs, not just a yes or no. Do not trust a single-day screenshot. SISTRIX tracked 82,619 prompts for 17 weeks and found that ChatGPT replaced 74% of cited domains week to week, AI Mode replaced 56%, and AI Overviews had substantial churn on a minority of prompts. Two runs will not erase that volatility. They will show you which observations deserve another look before you edit.

Score each engine separately. For a given engine, count the prompts that visibly cited your site and divide by the prompts you ran there. Then do the same within each intent group. Keep the ten short-versus-long pairs separate so they do not quietly change the weight of your main prompt set. Flag prompts that cited you on both baseline runs, on just one, or on neither. Those are different starting positions.

Expect some pages you rank well to be missing. An Ahrefs analysis of 863,000 keywords and 4 million AI Overview URLs found that 38% of cited pages were in the top 10, while 31% ranked beyond position 100. Google can split a prompt into sub-queries and cite sources for those, rather than simply cite the winner of the head term. The same study found YouTube accounted for 5.6% of AI Overview citations. Your ranking is useful context, not a substitute for the answer you can see.

At the end of week one, circle the prompts where competitors appear twice and you do not. Note whether the same competitor URL appears each time. That gives you a smaller, better-defined set of pages to investigate than “improve AI visibility.”

Weeks 2–3: edit five pages, and leave the rest alone

Choose up to five prompts where you were absent on both baseline runs and a competitor repeatedly appeared. Identify the page on your site that should answer each prompt. If you do not have one, mark the gap; do not pretend a vaguely related page is a control for an answer it never gave.

For the pages you do have, change one thing you can identify later: put a direct answer to the target question in the first paragraph. Keep a copy of the previous wording and record the edit date. Do not simultaneously rebuild the template, add a figure, change the title and rewrite the comparison. That would make any later movement hard to interpret. Leave other eligible pages untouched so you have something to compare against normal citation churn.

In the Princeton GEO benchmark across 10,000 queries, quotations lifted visibility 41%, original statistics 37%, and inline source citations 30%, while keyword stuffing did little. Those are results from that benchmark, not a promise for your five pages. They give you candidates for a later test if your own sheet points to an extraction problem. For this round, a clear answer near the top is enough of an intervention to track.

Do not expect the same edit to move every engine. One analysis reports that ChatGPT retrieves largely via Bing and averages about 7.9 citations per answer, Perplexity uses its Sonar index and averages about 21.9, and Gemini and AI Overviews use Google fan-out at about 13.3. It also reports roughly 82% overlap between Perplexity’s URLs and Google’s top 10, versus roughly 6.5% for ChatGPT. That is the reason to keep three result columns. One page edit can meet three different retrieval paths.

Re-run the target prompts in weeks two and three, recording the same fields as before. Resist the urge to make a second edit when one run disappoints you. You are trying to see whether the first change is followed by a repeatable difference, not conduct an emergency redesign between screenshots.

Week 4: read the result by engine, not by mood

Run all 30 to 50 prompts again on AI Overviews, ChatGPT and Perplexity. Use the same clean setup, location setting, prompt wording and scoring rules. Compare week four with both baseline runs, then look at the untouched prompts. If citations moved everywhere, your edited pages may not be the story.

Score citation rate per engine and intent group. Look at overlap, too: which prompts cite you across engines, and which cite you on only one? SISTRIX reports Jaccard domain overlap around 0.17 between AI Overviews and AI Mode and 0.125 between AI Mode and ChatGPT. Those are source findings, not results from your sheet. They are a good reason not to average your three engines into a single number. A page that gained on Perplexity and stayed flat on AI Overviews did not “half work.” It gained on Perplexity. Record that plainly.

Now turn the sheet into decisions:

  1. Cited after the edit, on repeated runs: keep the change and watch it next cycle. Do not credit the edit as the sole cause when the engine and untouched prompts also moved.
  2. Fetched near a run but never cited: inspect what the page actually exposes. Is the answer buried in an interactive block? Would a real HTML table make the comparison readable? A fetch narrows the investigation; it does not prove extraction failed.
  3. Cited on Perplexity but absent from AI Overviews: compare the cited page with the narrower questions Google may be answering. Do not assume one missing citation means the paragraph is bad.
  4. A competitor URL appears repeatedly: inspect that page, not the competitor’s whole site. Identify the question it answers and the structure that makes the answer easy to find. Test one corresponding change on your own page in the next cycle.
  5. No stable pattern: leave the page alone for now. More rewriting will not make a noisy sheet more informative.

That is the week-four output: pages to keep, pages to investigate and pages not to touch. It is a priority list, not a victory slide.

If the sheet wobbles, tighten the test

If week one and week four share almost nothing, first suspect volatility. Cut back to the ten prompts most directly tied to money. Run them at the same hour, on the same connection and with the same settings. Pair the citation log with downstream numbers. In GA4, I filter referrals for chatgpt.com, perplexity.ai and copilot.microsoft.com and compare them with citation dates. A citation that never sends a visit or an enquiry stays low on my fix list. A referral will not explain every answer, but it helps keep the priority list tied to what buyers do.

The short-list rule is simple: update pages that missed stable, high-intent prompts twice; create a page where the repeated answers cite a competitor and you have no relevant page; leave the rest until the next run. If you are also testing whether your pages get cited beside ads on commercial AI Overview queries, use the same ten prompts rather than inventing a second prompt set halfway through.

If you run a low-volume local book, say 15 searches a month for your core term, skip the full protocol. You may not get a stable read in 30 days. Run ten prompts once per engine, fix obvious access or page problems, and spend the remaining hours on reviews and service pages. For everyone else, keep the sheet and re-run it quarterly or after a visible engine change. Changing the prompts every time throws away your comparison.

Running this by hand once is useful. Running it continuously is the job groas takes on for clients through SEO and AI Search: specialized models check prompt-level citations across engines, investigate misses using fetch and page evidence, and ship page changes while a named strategist keeps the priority list tied to pipeline rather than citation counts.

Either way, the result changes what you do next. If citations repeat on prompts your buyers use, protect the pages earning them. If a competitor repeatedly gets the answer and you do not, test the relevant page. If the results keep moving, measure longer before you rewrite. Trust your sheet over anyone’s list of citation rules, mine included.