Search after the answer engine
Weigh the evidence for and against GEO, put a ceiling on what being cited is worth, and set up the three measures to read before you spend.
On this page
An answer engine reads the pages for the reader and replies in prose, with a few citations underneath. A page used to be ranked and then clicked; now it may be cited and not clicked, read by the engine rather than by the person. A new trade, "generative engine optimisation" (GEO) or "answer engine optimisation" (AEO), sells edits that make a page more likely to be cited. The evidence is contested, so the useful position is not a tactic but a ceiling on what the effort can return, and the measurements that would justify spending more.
Ranked, then cited
In classic search the reader saw ten ranked links and clicked one, and the click arrived in analytics as a session. An answer engine retrieves candidate pages, writes an answer from them and cites a few, delivering most of what it took from the pages on its own screen.
The best independent evidence is observational. Google users are less likely to click on links when an AI summary appears in the results (opens in a new tab) (Pew Research Center, 2025) tracked the browsing of 900 US adults in March 2025. On searches that showed an AI summary, people clicked a traditional result on 8% of visits, against 15% without one. They clicked a link inside the summary itself on 1% of visits, and ended their browsing on 26% of those pages against 16%.
Being cited is worth two things: the rare session, which analytics counts, and the impression, the company named in an answer the buyer trusted, which no click records. Sharp's How Brands Grow calls that mental availability; if it works, it shows up later as searches for the name, which is why measurement when the click disappears tracks brand search.
What GEO claims, and the evidence for it
The originating paper is GEO: Generative Engine Optimization (opens in a new tab) (Aggarwal et al., KDD 2024). The authors built a benchmark of 10,000 queries, rewrote source pages in several ways and measured each page's visibility in the generated answers. Adding citations, quotations and statistics to a page raised its visibility by up to 40% on their benchmark: quotations about 41%, statistics about 31%, cited sources about 28%. Keyword stuffing did worse than leaving the page alone. On 200 queries against a live commercial engine the gains were smaller, about 22% for quotations.
Read the metric before the number. "Visibility" in the paper is how much of the generated answer is attributed to the page, weighted by its position in the answer. It is not sessions, sign-ups or revenue. It is also the only peer-reviewed evidence for the tactics that we found; much of what else is published comes from agencies and tool vendors selling GEO.
The evidence against, and why it is your own measurement
The peer-reviewed evidence against is a benchmark too. C-SEO Bench: Does Conversational SEO Work? (opens in a new tab) (Puerto et al., NeurIPS 2025 Datasets and Benchmarks) re-ran the GEO paper's rewrites with four language models as the engine, across six domains, and found that most gained nothing significant; adding statistics lowered a page's citation rank in 19 of the 24 settings. Moving the page up the retrieved results, which is ordinary search ranking, did more than any rewrite, and the gains from rewriting shrank as more competing pages adopted it. Its outcome is citation rank, not sessions. The platform's own guidance is another piece: AI features and your website (opens in a new tab) (Google Search Central) says there are no additional requirements to appear in its AI features, "nor other special optimizations necessary", beyond ordinary search practice. That is an engine's owner describing its own engine, not a neutral party, and it says nothing about other engines. Note also that the paper's best tactics, citing sources, quoting people and adding numbers, are what a careful editor does anyway, so a retainer may be charging a premium for ordinary editing.
The rest of the case against is the reader's own data, and for the company below its first number is small: answer-engine referrals as a share of organic sessions.
Evidence check, October 2026: no peer-reviewed study testing GEO edits against a control on live sites, with sessions or revenue as the outcome, was found when this was written. This paragraph is re-sourced at each review of the course.
Worked: what a retainer could return
A company gets 20,000 organic sessions a month, and 120 of them arrive from answer-engine referrers: 0.6%. Twenty of its pages are already cited. An agency proposes a GEO retainer at $10,000 a quarter to rewrite them.
Put a ceiling on it first. If each of the 20 pages gains 30 sessions a month, the effort adds 600 sessions a month, 3% of organic. Value a session at $0.67, a third of a $2 paid search click:
ceiling 20 pages x 30 sessions -> 600 sessions a month
value 600 x $0.67 -> $400 a month, $1,200 a quarter
retainer -> $10,000 a quarter
break-even $10,000 / (600 x 3 months) -> $5.56 a session, about 8x an organic session
The retainer pays only if an answer-engine visitor is worth more than eight times an ordinary organic one, or if the uncounted part, citations that raise branded search, is large. Neither is known; each can be measured. So the decision is not "GEO works" or "GEO is a fad" but "we do not yet have the number that would justify $10,000".
What to measure now
AI-surface referrals. Sessions whose referrer is an answer engine, as a share of organic sessions: 120 of 20,000, 0.6%. Read it as a floor, because some answer-engine visits arrive with no referrer and count as direct.
Citation share. A fixed panel of buyer questions, asked of the answer engines every month, counting the share of answers that cite you and each named competitor. Answers vary from run to run, for the reasons in sampling and determinism, so ask each question several times. 50 questions × 5 runs is 250 answers; at a 12% share, 30 of 250, the standard error is √(0.12 × 0.88 ÷ 250) ≈ 2 points, and a month-on-month change has an error of √2 × 2 ≈ 2.8, so treat moves under about 5 points as noise.
Branded search volume. Searches for the company's name, where answers that name you without a click should show up. Google counts traffic from its AI features inside ordinary search traffic in Search Console's performance report (AI features and your website (opens in a new tab)), so the branded-query count there is the one to watch.
Then, and only then, a test: edit a randomly chosen half of the cited pages, leave the other half alone, and compare the halves on citation share and sessions over a quarter. A difference larger than the noise, with branded search moving, is evidence for this site; no difference is evidence too.
In practice
To the agency: "Which of the three numbers will move, by how much, and against which control pages?" A seller who answers with the 40% is quoting a benchmark's visibility metric, not your sessions. Offer a paid test on 10 of the 20 pages rather than a retainer on all of them, with the result that ends it written down first. The habit is the one from how to answer a case: name the number that would change your mind before spending.
To the board: "Answer engines send 0.6% of our organic sessions. We measure citation share and branded search monthly from now, and in a quarter we will know whether being cited is worth paying for."
The mistake that costs money is buying a retainer against a benchmark's visibility figure with no baseline of your own. $10,000 a quarter buys a ceiling of $1,200 of sessions, and without citation share and branded search measured beforehand, the rest of the claimed effect can be neither confirmed nor refuted.
Check yourself
1. Over two quarters, answer-engine referrals rose from 120 to 300 a month while organic sessions fell from 20,000 to 18,000. Do you move budget into GEO?
Not yet. 300 of 18,000 is 1.7%, and even doubling to 600 sessions is about $400 a month at $0.67. The 2,000 lost organic sessions, about $1,340 a month, matter more. Find out whether answers are taking clicks from pages you rank for: check citation share on those queries and branded search. Move budget once citation share is measured and the test on half the cited pages has read out.
2. After rewriting 10 pages, citation share on a 50-question, 5-run panel rose from 12% to 15%. Is the rewrite working?
Not shown. The difference between two months has a standard error of about 2.8 points, and 3 points is barely one of those. There is also no control: the engines change month to month. Compare the 10 rewritten pages with 10 untouched ones over the same months.
3. Answer-engine visitors might convert better than other organic visitors. With 120 such sessions a month at a 3% sign-up rate, how long before you could tell whether they are worth the $5.56 a session the retainer needs?
120 × 3% is 3.6 sign-ups a month, about 11 a quarter. That is enough to tell 3% from a rate eight times higher, but nowhere near enough to tell 3% from 4%. You will learn within a quarter whether the break-even is plausible, which is the question; precise conversion rates for this channel need more traffic than it has.