Last week a screenshot of my Search Console reached over 200,000 people on LinkedIn. It showed full AI prompts recorded as Google searches. Thousands of impressions. Zero clicks. The comments all asked the same thing: fine, the AI is reading my site, so how do I get into the answer?
I have been asked a version of that question in almost every sales call this year. Usually it arrives as "how do we get cited by ChatGPT". Sometimes it is "how do we increase AI citations". Occasionally it is a CMO who has read four playbooks, tried all of them, and wants to know why nothing moved.
So this is my attempt to answer it properly. Not with a list of tips lifted from a Google help page. With the exact framework we run on arobis.ai and on client sites, the numbers it produced, and the places where it did not work. We publish a page with this process in under two hours of machine time and about twenty minutes of human time, and I will show you every step :)
The short version
What is an AI citation, and why is it not the same as a recommendation?
An AI citation is a link or named source inside an AI answer. A recommendation is the AI naming your product as the thing to buy. Citations are the input. Recommendations are the outcome.
This distinction matters more than any tactic in this article, so let me be blunt about it.
When ChatGPT answers "what is the best CRM for a 40-person SaaS company", it does two separate things. It retrieves pages it trusts and reads them. Then it names companies. The pages it read are the citations. The companies it named are the recommendations. You can be cited without being recommended. Our own comparison posts are the proof: Google's AI quotes them constantly, and the brands named inside them get the recommendation, not us.
You can also be recommended without being cited, because the engine learned about you from a third-party page it read, not from yours. That is the whole reason we published the domains that get B2B SaaS cited in AI search. The citing page is often not your page.
So when I say "increase AI citations" in this article, I mean it as a step toward recommendation, not as the goal. If you only want the badge of being a source, most of what follows will still work. If you want pipeline, read how to build AI recommendation authority after this one, because that is where the citations get converted.
Where these numbers come from
Four sources, all dated, so you can judge them.
First, our own Search Console property for arobis.ai. We are an agency that publishes comparison content about AI visibility tools, and Google's AI leans on that content heavily, which makes our property a useful lab. All figures are from the three months to 18 September 2026 unless stated.
Second, a keyword and SERP pull run on 21 September 2026 in Semrush, with every ranking page's authority read at page level, not domain level.
Third, an engine test run the same day. One question, asked in a ChatGPT temporary chat with web search on, in Perplexity, and as a US Google search with personalisation off. I recorded every source each engine cited.
Fourth, two published studies I trust because they show their method: Rankscale's analysis of almost 8,000 AI citations across 57 queries, published on Search Engine Land in May 2025, and Adam Gnuse's audit of 15 domains and 7,500 ChatGPT referral sessions, published in November 2025. Where I quote a number from someone else, the link is right there.
Why most advice on AI citations does not move anything
Ten pages rank for the head terms. They average about 2,500 words, most cite the same three studies, and none shows its own dated results or separates being cited from being recommended.
Before writing this I read every page that ranks in the US top ten for "how to increase AI citations" and "AI citation optimization". That is the fourth strategy below, applied to this exact article, so you can see what it produces.
Here is the grid. The last column is the one every single page fails on.
Two things jump out. Only three of the ten bring any data of their own, and two of those three are on the same site. And the pieces with real data are the ones Google's AI Overview actually cites for this question. That is not a coincidence, and it is the argument for strategy five.
The other thing I want you to notice is what is missing. Meltwater's 73% lift is real and I respect them for publishing it, but it counts citations without saying whether a single extra buyer chose Meltwater. That gap, cited versus recommended, is where the money is, and nobody on the list touches it.
How to increase AI citations: the 8-strategy framework we actually run
The framework is eight steps in a fixed order. Six of them happen before writing. Skipping the first six is why most content programmes fail without ever finding out why.
Here is the whole thing on one screen. Each strategy gets its own section below with the numbers behind it.
Strategy 1: Start from the demand you already have, and cut the machine noise
Your Search Console already lists the queries you nearly rank for. Sort by impressions, drop everything over 55 characters, and the cheap wins are what remains.
Most teams pick topics from a keyword tool. We pick them from Search Console, because Search Console knows what Google is already tempted to show us for. Sort queries by impressions, not clicks. Clicks hide the zero-click terms, and zero-click terms are exactly where AI answers live.
Then apply the rule that changed how we read the whole report. Drop every query over 55 characters. Those are not people. They are the hidden sub-questions Google's AI generates when it splits a prompt into pieces, a process Google itself calls query fan-out in its guidance on AI features in Search. Monitoring tools add more of them. They inflate impressions and never click.
We measured it on our own property. In a 1,000-row export, the short queries carried all of the clicks and the long ones carried none.
The third row is this article's own demand pool. Sixty-one queries about being cited, average position 8.8, and not one click, because 58 of the 61 are long AI prompts like "what is a practical 90 day roadmap to start generative engine optimization and test what gets cited". That row told me two things: Google's AI already associates arobis.ai with this topic, and we had no page built to receive it. The full breakdown of what those prompts look like is in our study of 178 strange Search Console queries.
The output of this step is a ranked list of clusters where you already have impressions and sit between position 8 and 30. It is a table, not a plan. Recommendations come after step two.
Strategy 2: Prove you can win the specific pages, not the keyword difficulty score
A difficulty score describes the average site. The real test is whether a page with your authority can displace the ten specific pages ranking today. Read their page-level authority, then decide.
This is where half of every shortlist dies, and that is the step working.
For this article the obvious head term was "AI citation". Semrush shows it at 390 searches a month in the US, and I nearly chased it. Then I read the ranking pages. APA Style. Purdue's library. Scribbr. The term means "how do I cite ChatGPT in my essay". A marketing article would never rank there and would deserve not to.
The term I chose instead, "how to increase AI citations", has 140 searches a month, and the ten pages ranking for it have page-level authority scores of 14, 20, 50, 9, 0, 9, 22, 10, 49 and 4. Our domain sits at 16. That means we out-authority six of the ten before we write a word. The same check on "AI citation optimization" gave 26, 10, 8, 20, 0, 8, 0, 11, 0 and 11. Both winnable. Both under the difficulty ceiling we set for ourselves.
Notice what I compared. Not our domain score against a keyword difficulty number. Our domain score against each ranking page's own score. A page on a strong domain can still be weak, and a Reddit thread at zero can still hold position one because Google wants a forum in the mix.
The rule I give the team is simple. Better to kill a term now than after 4,000 words.
Strategy 3: One page per intent, and count the overlap before you add another
Two of your own pages chasing the same intent split the signal and both lose. Crawl your sitemap, count the head-term phrase on every page, and retitle or merge before publishing.
Cannibalization is invisible unless you count. So we count. For this article I crawled all 108 URLs on arobis.ai and scored every one for how often it uses "cited" or "citation" and whether the phrase sits in a title or heading.
The result surprised me. We had nine pages that mention citations more than 30 times each, including our Perplexity guide, whose title literally promises to explain what gets brands cited. But none of them targets the topic as its primary term. They are engine-specific spokes. This article is the hub they were missing.
The crawl also caught something a human reads past. Two of our posts argue that being cited is the wrong goal. They are right, and this article agrees with them, which is why the first section above defines cited against recommended before saying anything else. If I had not run the crawl, this page would have quietly contradicted two others on the same site, and AI engines notice contradictions across a domain far more reliably than readers do.
The verdict for every candidate is one of three: safe to publish, retitle an existing page first, or update the existing page instead. Write it down. It is the cheapest decision in the whole process and the one most often skipped.
Strategy 4: Read all ten ranking pages and find the column they all fail
You cannot out-write pages nobody on your team has read. Put all ten in one grid with identical columns and the shared gap becomes visible. That gap is your angle.
You saw the grid for this topic earlier. I want to explain why the grid works when a summary does not.
When you read ten competitor articles one after another, you come away with a feeling. "They are all a bit generic." A feeling is not a brief. When you force each page into the same seven columns, you get a fact: zero of ten separate cited from recommended, seven of ten have no data of their own, and the three that do are the ones the AI Overview cites. Now the brief writes itself. Bring dated data. Draw the line between cited and recommended. Keep the table and the FAQ because the two strongest pages have one each and none has both.
The columns I use every time are word count, number of items or steps, stated methodology, every H2, hard numbers published, any original research, author credentials, last updated date, and the three biggest weaknesses. Do not soften the weaknesses. The point is to find the thing all ten are missing, and the answer is usually the same: original data.
Strategy 5: Ship one number nobody else has
Anyone can generate 5,000 words. A reproducible measurement with a date and a method cannot be copied, and it is the single strongest predictor of being cited across every study I trust.
This is the most important strategy in the framework and the one almost everyone skips because it feels like extra work. It is twenty minutes.
The evidence for it is unusually consistent. In the Search Engine Land audit of 15 domains, 52.2% of blog posts that earned ChatGPT citations contained original or owned data, and posts combining a clear answer with owned data were the strongest configuration in the set. The Princeton GEO paper, presented at KDD 2024, reported visibility gains of up to 40% in generative engine responses from optimisations that included adding statistics and citing sources. And Surfer's own analysis found that 67.82% of the sources AI engines cite do not rank in Google's top ten, which tells you engines are selecting for the passage and the number, not the position.
The cheapest original data you can produce is the engine test. Here is exactly how we run it, and what it returned for this article.
Write three prompts a real buyer would type. Full sentences with context, not keywords. Run each one on ChatGPT in a temporary chat with web search on, on Perplexity, and on Google with personalisation off. A logged-in ChatGPT account gives you a personalised answer, not what your buyers see, so the temporary chat is not optional. Record the companies named, in what order, and every source cited. Then publish two tables: companies by how many runs named them, and sources by how many citations they earned. State the date, the location, the run count and the limits of the sample.
Here is what the three engines cited on 21 September 2026 when I asked how a B2B SaaS company gets its content cited by AI search.
Three engines, three almost non-overlapping source pools, and not one of the ten pages from the teardown grid appeared in ChatGPT or Perplexity. If you only optimise for the Google result, you are optimising for one third of the answer.
This matches the larger pattern Rankscale found across 8,000 citations, and their table is worth keeping next to yours.
Source: Rankscale data as published on Search Engine Land, May 2025, 57 queries run repeatedly across four engines. The B2B split in the same study is the line I quote most to clients: for B2B queries, company sites and vendor blogs made up about 17% of all citations. Your own blog is a legitimate citation source in B2B. It almost never is in consumer categories.
If you have a product with usage data, you have a second option that beats the engine test. Anonymise an export, find the three most surprising and defensible findings a buyer would not already know, and publish each one as one sentence with the number, the sample size, the period, and what it does not prove. Our AI Readiness Report is that method applied to 1,188 websites that ran our checker. The finding that 19.8% of sites block GPTBot in practice while only 4.8% mean to has been quoted back to me by people who never visited the report page. That is what a citable number does.
Strategy 6: Fact-check every claim at its source, with today's date
Competitor articles are wrong constantly and the errors spread through AI answers. Read every company's own site today, quote only what it states, and write "not published" where it says nothing.
This step is boring and it is where a technical buyer decides whether to trust you.
The rule is strict. Never take a figure from a listicle or a third-party article. If a price is not on the vendor's own site, write "not published". Do not estimate. Do not infer. Record the source URL and the date for every claim.
We learned this the expensive way. In September we re-verified the pricing in our tool comparison posts by reading the vendors' pages in a real browser, because their prices render with JavaScript and a plain fetch sees nothing. Three of our live pages were wrong: one vendor's starter plan was $95 a month, not the $80 we had copied from an older article, another's entry plan was $250, not $300, and a third had launched a free tier we did not mention. Every one of those errors had been copied from a competitor's listicle that was itself out of date.
Here is why it matters for citations specifically. Engines cross-check claims across sources. When your number disagrees with the vendor's own page, you are the one who gets dropped. When your number agrees and carries a date, you become the convenient source for that fact. The "not published" gaps are themselves worth writing about, because a buyer searching for a price that no vendor publishes will land on the page that says so honestly.
Strategy 7: Write for extraction, not for engagement
Under every question heading, write a direct answer of 20 to 25 words with no links in it, lead with your own data, and add a methodology section and an FAQ.
Now, and only now, write. Quality is set by the constraints you give before the first word, not by editing afterwards. These are ours.
The answer capsule. Directly under every heading that is a question, or that implies one, write a self-contained answer of about 120 to 150 characters. Adam Gnuse's audit found that 72.4% of blog posts receiving ChatGPT citations had one, and that around 91% of those capsules contained no links at all. Links inside the capsule tell the model the real answer lives somewhere else. Put your links in the paragraphs below it. Every H2 in this article follows that rule, and you can check it.
Lead with your own data, high on the page. The engine test table should be in the first third, not the last. Engines weight the opening of a page. So do readers who are deciding whether you know anything.
The target term in the title, the H1 and one H2. Once each. Spelled out in full. We had a listicle sitting at position 73 for "best answer engine optimization agencies" with 7,382 words that used the acronym everywhere and the full phrase twice. Retitling it was a ten-minute fix.
A methodology section. What you tested, when, the limits, and any conflict of interest. Mine is the "where these numbers come from" section above. It is the section engines cite when they want to justify trusting a page.
A "who this is not for" line on every recommendation. If you rank tools or agencies, say who each one is wrong for. Honest exclusions are the strongest credibility signal a vendor-authored page can send, and Rankscale's data shows vendor pages get cited in B2B precisely when they read as fair.
Disclose in the first screen. If you appear in your own ranking, or you sell the service you are describing, say so at the top. I did it in the first paragraph of this article. It costs nothing and it removes the one objection every sceptical reader and every retrieval model is looking for.
A "Frequently asked questions" heading with buyer-phrased questions. Eight of them, each with a two or three sentence answer that stands alone. On our site that heading triggers FAQ schema automatically, so the questions become structured data without any extra work. Check your own template does the same, or add the markup by hand.
Mechanics. One internal link per destination, with varied anchors. No em dashes, because they read as machine-written and buyers are tired of it. Brand names written in full. Only the HTML tags your CMS actually renders, because a table it strips becomes a blank gap on the live page.
One more thing on structure, since the per-engine differences are real. Perplexity wants the answer inside the first 150 words and punishes stale pages. ChatGPT cares about entity strength and authoritative reference sources. Google's AI wants you ranking somewhere in the top 20 for the cluster and then picks the passage. We keep a guide for each: ChatGPT, Google AI Overviews, Gemini, Claude and Microsoft Copilot. The eight strategies here are what those guides share.
Strategy 8: Ship it, link it, request indexing, then measure at fixed checkpoints
Add fifteen internal links from existing posts, request indexing yourself, and judge the page at two, four and eight weeks, never at week one.
Publishing is not the end of the process. It is the start of the part most teams never do.
Verify the live URL, not the draft. Your CMS, your template and your scripts all get a vote in what renders. Crawl the live page. Every link returns 200. Title and meta description present. Canonical appears once, on the preferred domain. Tables in the right positions. Schema present, checked in a real browser, because script-injected schema is invisible to a plain fetch.
Internal links. Find fifteen existing posts where a link to the new page genuinely helps the reader. Skip the posts the new page already links out to. For each one, record an exact quote from a single paragraph to search for, the full sentence to add after it, and the exact words to hyperlink. Different anchor text for all fifteen. This is the cheapest authority you fully control. On our site, the only pages that broke onto US page one this year were pages we had linked from at least a dozen existing posts first.
Request indexing. Google reserves this click for humans. Open Search Console, paste the URL into the inspection bar, wait for the check, click Request Indexing. Do it for every existing post you edited too. Skipping it can cost weeks of waiting for a natural crawl. Sixty seconds, every page, every time.
Then wait, on a schedule. Judging a page on week-one traffic kills good pages and keeps bad ones. These are our checkpoints.
The eight-week engine re-run is the checkpoint that matters for citations, because it is the only one that measures the thing you set out to change. Search Console's Generative AI report, which rolled out to every property on 31 August 2026, tells you AI impressions by page but not queries or clicks. The engine test tells you which sources won. You need both, and I explain how to wire the tracking in how to track AI traffic to your website.
What this framework produced on our own site
Our checker page went from US position 80 to page one in a day and was cited inside the AI Overview above it. AI features now make up 19% of our impressions.
I promised a dated before and after, since none of the ranking pages has one. Here is ours, with the caveats.
Our free AI Visibility Checker page had been sitting around position 80 in the US for its main term. We ran the eight strategies on it: rebuilt the demand pool from Search Console, confirmed the ranking pages were beatable on page-level authority from position four down, added the explainer sections and FAQ that the teardown showed every competitor had, published a proof strip with our own scan numbers, and linked it from every relevant post on the site. On 29 August 2026 it moved to a US position around 13 in Search Console, which is the bottom of page one once you account for the AI Overview block above the organic results. It was also cited inside that AI Overview. US impressions went from 40 to 90 a day to 210 to 240 a day, and we had five inbound leads that week, three from the US.
Across the whole site, the Generative AI report in Search Console shows 22,100 AI impressions out of 117,000 total for the 28 days to 17 September 2026. That is 19%. Our comparison posts sit at 22 to 27% AI share each. The checker page sits at 13% AI share and gets most of the clicks. Which brings me back to where this article started: the pages the AI quotes most are not the pages people click, and being cited is a step, not the finish.
On the client side, a B2B marketplace we work with now sees 16% of its Google impressions inside AI features, and its best comparison listicle sits at 52% AI share. Same eight steps. Their numbers are theirs to publish, so I will leave it at the share figures.
The caveats, because a number without them is marketing. This is one site and one client over one summer. The checker page had an existing backlink profile at its old URL that we redirected. The 19% figure includes brand queries, where our homepage sits at 57% AI share. None of this proves causation. It shows the sequence and the timing, and you can reproduce both.
When this does not work, and what to do instead
The framework fails when the demand pool is empty, when every ranking page out-authorities you by 30 points or more, or when your category is answered by primary documentation rather than blogs.
I want to be honest about the limits, because the ranking pages are not.
It does not work when Search Console shows nothing. If your site has no impressions in a topic, strategy one returns an empty table, and you are building demand from zero. That is possible, but it is a different playbook with a different timeline, and the first six months look like failure.
It does not work when the authority gap is too wide. If every page in the top ten sits 30 or more points above you at page level, the honest verdict is walk away, or earn the mention on a page that already ranks instead of building your own. That is what our study of the domains AI cites is for.
It does not work for questions engines answer from documentation. Look at the ChatGPT row in the engine table again. For "how do I get cited", ChatGPT went to OpenAI's and Google's own docs and ignored every blog. If your buyers' questions are the kind an engine answers from a manual, your best move is to become the manual: publish the specification, the definitions, the reference tables that documentation would contain.
And it does not turn citations into pipeline on its own. That needs the recommendation layer, which is a different set of work on third-party pages, reviews and comparisons. We map it in the AI visibility audit before any content is written, and it is the reason our engagements on the pricing page are scoped as demand generation rather than content production.
Run the engine test on your own site today
You can run strategy five on your own domain in about 25 seconds with our free checker, then compare what the engines actually cite in your category against your own pages.
Open the free AI Visibility Checker, enter your domain, and it will fetch your homepage as GPTBot, ClaudeBot, PerplexityBot and Googlebot and score six signal groups. That is the crawl access part of strategy seven done. If you want the ChatGPT side specifically, the ChatGPT Visibility Checker runs your buyer prompts and shows whether your brand is named.
Then do the manual part. Three buyer prompts, three engines, one spreadsheet. Publish the source table with the date. You will have one number nobody else in your category has, and that is the whole point.
Frequently asked questions
How long does it take to increase AI citations?
Two to eight weeks for a page that already ranks in the top 20 for its cluster, based on our checkpoints. Perplexity can cite a new page within days because it retrieves live. Google AI Overviews follow the organic index, so they move when your rankings move. ChatGPT is the slowest and depends most on third-party references to your brand.
What is the difference between an AI citation and an AI recommendation?
A citation is your page being used as a source in an answer. A recommendation is the engine naming your product as the one to choose. You can have either without the other. Citations are the input you control most directly; recommendations are the outcome that produces pipeline.
Which AI engine is easiest to get cited by?
Perplexity, in our tests, because it retrieves live pages and rewards narrow, specific framing regardless of domain size. Google AI Overviews are next if you already rank in the top 20. ChatGPT is hardest, since it leans on Wikipedia, major news and primary documentation and cites vendor blogs only about 1 to 3% of the time.
Does word count matter for AI citations?
Not directly. The audit of 15 domains found the answer capsule and original data predicted citations, not length. Length matters only because a page that covers the fan-out sub-questions tends to be longer. Write what the sub-questions need and stop.
Should I put links inside my answer capsule?
No. Around 91% of the capsules in cited posts contained no links at all. Keep the capsule self-contained and put internal and external links in the supporting paragraphs directly below it.
Do I need schema markup to get cited by AI?
It helps and it is cheap, so yes. Article schema with a named author, Organization schema, and FAQ schema on your question section are the three that matter. They do not replace the content signals in strategy seven, and no study I have seen shows schema alone earning a citation.
How do I measure AI citations without a paid tool?
Three ways. Search Console's Generative AI report shows AI impressions by page. A GA4 channel group for AI referrers shows visits from ChatGPT, Perplexity and the others. The manual engine test in strategy five shows which sources won for your prompts. Together they cover impressions, clicks and citations.
Can my own blog be a citation source, or only third-party sites?
In B2B, yes. Rankscale's study found company sites and vendor blogs made up about 17% of citations for B2B queries, mostly comprehensive comparison pages that cover competitors fairly. In consumer categories the share drops below 4%, so third-party placement matters far more there.
