Your buyers have started asking an assistant which product to use, and the answer names three companies. If yours is not one of them, you never see the visit, the search impression or the bounce. Nothing shows up in your analytics at all. That is why how to get cited by ChatGPT has become a real budget line and not a curiosity, and why most advice about it is so unsatisfying: it is written as a checklist of tactics with no way to tell whether any of them worked. This piece covers the mechanism, the five things that plausibly move citations, and the measurement that separates a change you caused from drift you did not.
Why how to get cited by ChatGPT is a measurement problem first
An assistant's answer is not a ranking. Ask the same question twice and you can get two different sets of names. Ask it from a different account, or a week later, and it moves again. There is no rank tracker that makes this stable, because the instability is in the thing itself, not in your instrument.
That has two consequences, and both are practical.
The first: a single check tells you almost nothing. One answer that names you is a sample of one from a distribution you have not observed. You need the same question asked many times before "we get mentioned" becomes a number rather than an anecdote.
The second: a change over time is not evidence of your work. These platforms update on their own schedule. Something you did in week one and something they shipped in week three both show up as the same movement in the same chart. Untangling them is the entire job, and it is why Rankli runs citation tracking as an experiment with a control group rather than as a dashboard.
Where a citation actually comes from
Two paths, and they respond to different work.
Retrieval
The assistant runs a search at answer time, reads a handful of pages and writes from them. Being cited here means being one of the pages retrieved and being useful enough in that moment to survive into the summary. This path is fast-moving and it is the one that ordinary publishing can affect within weeks. It also rewards a specific shape of page: one that answers a named question directly, near the top, in language that matches how the question was asked.
Training
The model answers from what it absorbed before the conversation started. Being named here is a slower business, and it is downstream of how often your product is described on pages that other people wrote: comparison posts, directories, forums, roundups, documentation that quotes you. You cannot edit that corpus, but you can be worth including in it, and you can make the facts about you consistent everywhere they appear.
Most answers mix the two. The practical reading is: publishing moves the retrieval half quickly, and the training half follows the reputation the retrieval half builds.
The five things that move AI citations
Answer the exact question, high up the page
Assistants summarise. A page that spends 400 words warming up gives the summariser nothing to lift. Put the direct answer in the first paragraph under the heading that matches the question, then earn the rest of the read. This is also the structure our scorer enforces: the target phrase has to appear in the title, in the first 100 words and in at least one H2, which is a crude proxy for exactly this shape.
Be present where the comparisons happen
When someone asks an assistant for the best tool in your category, the retrieved pages are usually roundups, alternatives lists and comparison posts, mostly written by other people. Being nameable there is a content job: publish the comparison page for your own category, keep every competitor figure sourced and dated, and be the page that is worth citing rather than the page that begs to be. Our own comparison hub is built that way, with each figure read off the vendor's live page on a stated date.
Let the crawlers in
Nothing gets retrieved that cannot be fetched. Check your robots.txt against the assistant crawlers by name, check that your key pages are server-rendered rather than assembled in the browser, and check that nothing important sits behind an interstitial. Rankli's own crawler behaviour is documented on our bot page, and the same discipline applies to whoever is fetching you.
Keep your facts identical everywhere
If your pricing page says $99 and your homepage implies $79, a model that absorbed both will state one of them at random, and often the wrong one. Pick the canonical number, put it in one place in code if you can, and render it everywhere from there. Consistency is not a growth tactic. It is what keeps a machine from inventing a version of you.
Publish at a cadence, on the questions buyers actually ask
Citation share follows coverage: you get named on questions you have written something useful about. Thirty articles a month on the questions your buyers ask is a different asset from thirty articles a month on whatever had search volume. This is the whole reason Rankli plans from your keywords and your category rather than from a generic list, and why every draft has to mention what you actually sell to pass its checks.
How to measure it instead of guessing
Pick the questions, then randomise
Start from the commercial questions a buyer would type: "best X for Y", "X vs Z", "is X worth it". Rankli discovers them from your category and your competitors, clusters them, and splits them into arms before any measuring happens. The split has to come first, because a control group chosen after you have seen the results is not a control group.
Ask each question many times, on every engine
A single answer is noise. Rankli asks each tracked question repeatedly on ChatGPT, Perplexity, Gemini and Claude, then reads each answer for which brands appear, in what order, and whether the answer absorbed the point without naming anyone. On the Growth plan that is 240 tracked questions with six runs each across four engines, which is enough repetition to put an error bar on the number rather than a single reading.
Hold out a control and pre-register the prediction
Before the work starts, Rankli writes down the intervention, the questions it should move, the size of the effect expected and the questions deliberately left alone. Then it re-measures and compares the treated set against the held-out set. The comparison is a difference-in-differences: the change in your questions, minus the change in the questions you did not touch, which is where platform drift cancels out.
Read the difference, not the line
Both numbers are shown, and they usually disagree. That is the point. Our calibration harness builds 150 independent worlds and runs the same follow-up window twice, once with the intervention and once without, from identical seeds, so the true causal effect is known exactly. Against a true effect of 2.3 points, the difference-in-differences estimator's interval covered the truth 92.7% of the time with a bias of −0.03 points. The naive before-and-after estimator overstated the same effect by 31%, and in windows where nothing was done at all it still moved at least 1.3 points half the time. Full method and figures are on the proof page.
The pages that get named most often
Across the questions we track for our own category, the pages that show up inside answers have a family resemblance, and it is not a subtle one.
They are specific. A page called "Outrank alternatives" with nine products, each with a price read off a live vendor page and a date attached, gives a summariser something to quote. A page called "the future of content marketing" gives it nothing.
They are structured for extraction. One question per heading, the answer directly underneath, lists where a list is genuinely the shape of the answer. Assistants lift paragraphs, not vibes.
They carry numbers that can be checked. Prices, limits, counts, dates. A page that says "affordable" competes with everyone; a page that says "$99 a month for 30 articles, read on 4 September 2026" is the one worth citing.
They state who is speaking and when. A visible publication date, a named organisation, and figures that agree with the rest of your site. This is also what our own scorer is testing on every draft, from the other direction: an article with an unsourced statistic in it fails a hard rule, which means it never becomes one of these pages in the first place.
None of that requires writing for machines. It is what a well-made reference page has always looked like. The change is that a badly made one now loses in a place you cannot see.
A four-week routine that fits in a normal week
Week one: fix crawlability and consistency. One canonical price, one canonical product description, crawlers allowed, key pages server-rendered.
Week two: publish the two pages that answer your category's most-asked buying questions directly, with the answer in the first paragraph.
Week three: publish your comparison pages, sourced and dated, and get the facts about you correct on any third-party listing you control.
Week four: read the measurement. Not the raw before-and-after, the controlled difference, and be willing to conclude that a change did nothing.
Then repeat. Citation share moves slowly, in a market where every competitor is also publishing, and the compounding comes from doing this every month rather than from any single clever move.
What to avoid
Do not buy a tool that reports a single citation score with no error bar and no control. It will move on its own and you will make decisions on the movement. Do not write pages aimed at models rather than people, because the summariser is reading the same words your buyer is. And do not treat one flattering answer as a result: ask the same question ten times before you believe it.
Rankli measures citations across all four engines on every plan, including the $1 trial: three finished articles published to your own site, plus a measurement round, in fourteen days. Cancel from the app any day, unused articles refunded. Plans and limits are on the pricing page.