I use AI to find where your revenue leaks, get you named inside ChatGPT, Perplexity and Google AI Overviews, then convert that attention into monthly cash flow with email.
Generative engine optimization (GEO) is the practice of getting a brand named inside AI-generated answers — the paragraph ChatGPT, Perplexity, Google AI Overviews or Copilot returns instead of a list of links. It is a different discipline from search engine optimization: it turns on entity consistency, extractable answer formats and third-party corroboration rather than rankings alone.
I run GEO — answer engine optimization, if your team calls it that — alongside technical and content SEO, and lifecycle and outbound email, as one program, because AI-referred traffic is high-intent and low-volume — it converts at roughly four times the rate of ordinary organic traffic, and it leaks without a sequence behind it. Every engagement starts with a revenue baseline and a prompt set of the real questions your buyers ask, then reports against both monthly.
Typical clients are direct-to-consumer brands, B2B SaaS companies and multi-location service businesses doing $50K to $1M a month, where a shortlist decision is being made by an assistant before anyone visits a website. If your category is one an assistant can answer in a paragraph, you are already in this fight whether or not you are competing in it.
The short version
None of these are mine. They are from independent studies you can open and check — the same standard I hold my own reporting to.
Position one in Google buys you almost nothing here. The assistant retrieves from a different pool than the one you have been optimising for.
Semrush, AI Search & SEO Traffic Study, July 2025
Measured by conversion rate. AI-referred visitors arrive further down the funnel because the assistant already did the shortlisting.
Semrush, AI Search & SEO Traffic Study, July 2025
Across 300,000 keywords, comparing March 2024 with March 2025. Your rankings can hold perfectly still while the traffic behind them drains.
Ahrefs, AI Overviews Reduce Clicks, 2025
Software returns 36:1, agencies 42:1. This is why every engagement ends in email — it is the highest-return place to put AI-referred attention.
Litmus, The ROI of Email Marketing, 2025
Where the two disciplines actually compete
Searches showing an AI Overview
Of those, the share that were purely informational
Same website, two different games. This is the comparison I draw on a whiteboard in the first call, so it may as well be on the page. The longer version, with the evidence →
| Dimension | SEO | GEO / AEO | Why it differs |
|---|---|---|---|
| What you win | A slot in a list of ten links | A named mention inside one generated paragraph | There is no slot four in an answer |
| Unit of measurement | Keyword ranking position | Citation share across a prompt set | Answers vary per user; a single rank has no meaning |
| Where the source sits | Positions 1–10 | ~90% of ChatGPT citations come from 21+ | Retrieval samples far deeper than a results page |
| Primary lever | Links, keyword coverage, crawl health | Entity consistency, extractable answers, corroboration | Models resolve who you are before deciding to name you |
| Content shape | Long-form built to hold a ranking | Direct answer in the first 100 words, then evidence | Retrievers lift passages, not pages |
| Structured data | Optional; earns rich results | Load-bearing; it is how the claim is machine-read | Schema is the least ambiguous statement on your site |
| Third-party pages | Useful for links | Frequently the cited source instead of you | Assistants prefer roundups and review platforms |
| Traffic profile | Higher volume, mixed intent | Lower volume, ~4.4× the value per visit | The buyer arrives already qualified by the assistant |
| Time to signal | 3–6 months for competitive terms | 6–10 weeks for entity fixes, a quarter for share | Entity corrections propagate faster than authority |
They are not alternatives. Technical health, crawlability and clean information architecture are shared inputs — a page an assistant cannot parse is almost always a page a crawler handles badly too. The mistake is not doing SEO; it is assuming that finishing SEO finishes the job. Ahrefs measured a 34.5% clickthrough drop for top-ranking pages once an AI Overview appears above them, which means the same ranking now returns roughly two-thirds of what it used to.
What changes is the order of operations. In an SEO program, content volume comes early and structure comes late. In a GEO program it inverts: the entity gets fixed first, then schema, then a small number of genuinely extractable pages, then outreach to the third-party pages the models already retrieve from. Volume is the last lever, not the first, and in most engagements it is never pulled at all.
Pick a buyer question, then flip between what an assistant said before the work and after it. Same question, same engine — one of them names the client. Or put your own brand in it at the bottom.
Nothing is sent anywhere — this runs in your browser and just shows you what being the recommendation looks like.
Modelled engagements, each built on the published benchmarks above. Every one names the research it is calibrated against — open a case and you will see it.
Every engagement runs the same sequence. The AI work is not a content shortcut — it is how we read the customer data fast enough to know which of these five actually holds the revenue.
Support tickets, reviews, won and lost CRM notes, order intervals — clustered by a language model into the questions buyers actually ask and the segments actually worth emailing.
Not keywords — the 150–400 real questions your buyer types into an assistant, weighted by the revenue behind each. That becomes the scoreboard.
One name, one category, one claim everywhere. Then answers written to be extractable: the direct answer in the first 100 words, real numbers, comparison tables.
Models trust consensus, and they pull heavily from pages ranking well below your own. Get named in the listicles, comparison pages and review sites they retrieve from.
AI search traffic is high-intent and low-volume. Lifecycle and outbound sequences are what turn it into cash collected this month instead of a traffic chart.
Almost every guide in this field stops at “get cited.” The best available data says that is roughly half the job, and the half that is left is the one that decides revenue. Full breakdown, and the five fixes →
An assistant can quote your page and still recommend your competitor. Ahrefs ran 34 self-promotional pages across five domains and read 9,886 AI answers from ChatGPT, Gemini, Perplexity and Copilot between February and May 2026. Among the answers that cited their page, 43% never mentioned their brand at all — the model used their research and named someone else.
The assistant read the source and recommended a competitor out of it instead. Citation was not the win it looked like.
Only 6% came from answers that cited its own content. You cannot self-describe your way into an answer.
About one day in three. Checking your visibility once and drawing a conclusion is close to guessing.
Same pages, same brand, same window. Which questions you go after matters more than how many.
Ahrefs, Self-Promotional Content and AI Search, 2026. 34 pages, 5 domains, 9,886 answers, four engines, February–May 2026.
What this changes about the work. If citation were the finish line, the job would be a content brief. It is not, so four things get built that a pure citation strategy skips: the cited page has to state your differentiator in the same paragraph as the category term, so a model lifting that passage cannot separate them; the third-party pages have to carry it too, because that is where 94% of new mentions came from; the prompt set has to be narrow enough to win, since a 66.4% mention rate on a specific question beats 15.8% on a vague one; and measurement has to run daily, because a page that shows on one day in three cannot be assessed by looking once.
And it changes what gets reported. Citation share alone would have called that Ahrefs experiment a success. The scoreboard here tracks two numbers per prompt: whether you were cited, and whether you were recommended — named as the answer rather than used as a footnote. The second number is the one joined to the CRM, because it is the one that correlates with pipeline. Where the gap between them is wide, the fix is almost never more content; it is that the page an assistant lands on is doing a better job of selling somebody else.
The strategy is mine. The builds are not a one-person job, and pretending otherwise would show up in the delivery dates. Engineering runs out of Lahore, client and commercial work out of Celebration, and the Gulf and EMEA accounts out of Abu Dhabi.
Diagnosis, prompt-set design, entity strategy and every readout. This is where the audit gets written and where the monthly revenue review happens. US and Canadian accounts run from here.
Where the software gets built: Citegeist’s tracking layer, the Revenue CRM, the custom systems and every AI-native site. Four engineers and a data engineer, with an overlap window that covers both other offices.
Regional accounts, Arabic-language prompt sets and the multi-market entity work that comes with them. An assistant answering in Arabic retrieves from a different pool than the same question in English, and that has to be measured separately.
How the work is actually split. Every engagement has one strategist accountable for the number and one engineer accountable for what ships. You are never handed to an account manager who relays questions. The prompt-set design, the entity decisions and the monthly readout are mine; the tracking infrastructure, the schema compiler, the CRM joins and the site builds are the engineering team’s. When a build is in flight you get the engineer in the call, not a summary of what they said.
Why three offices and not one. AI visibility is not one market. A brand can be named confidently in English and be invisible in Arabic, because the retrieval pool is different, the corroborating sites are different and the category language does not translate cleanly. Running a Gulf office means those prompt sets are written by someone who reads the answers natively rather than through a translation layer. The Lahore team exists because the software is the product now, and a five-person engineering team is the difference between shipping tools and describing them.
Twenty buyer questions. Four answer engines. One brand, and the honest reason the assistants name it or walk past it. No theory, no predictions about where search is heading — just what four machines actually said this week, and what I would change on Monday. Roughly half the teardowns end with the brand not needing me, and those are the ones people forward.
There is a particular kind of losing that looks like winning. Your page gets cited. The assistant quotes your research, borrows your framing, uses your numbers — and then recommends somebody else. Ahrefs measured it across 9,886 answers: 43% of the ones citing their own page never named their brand at all.
This week I pull apart a brand doing exactly that. Three pages that get quoted constantly, the competitor being recommended straight out of them, and the small structural habit causing it — a claim and a category sitting one paragraph apart, which is all the distance a model needs to take one from you and the other from someone else. The rewrite is in there, before and after, with the ratio re-measured a fortnight on.
The most common finding in every audit I run, and the one that takes ten minutes to fix. Somewhere between your CDN, your security plugin and a default nobody chose, GPTBot and PerplexityBot are being turned away at the door — usually filed alongside the scrapers, usually by a checkbox somebody ticked in 2023 to stop content theft.
Nothing downstream of that matters. Not the schema, not the rewrites, not the outreach. The page is never read. So this issue is the ten-minute check, done four ways — Cloudflare, Hostinger, Vercel, AWS — with what each provider’s defaults quietly do, and how to let the retrieval crawlers in while keeping the training ones out, if that is where you want the line.
Google-Extended does almost nothing people think it doesOne email a week. The teardown, the numbers under it, and nothing else — no drip sequence waiting behind the signup, no webinar invitation, no second email on Thursday. Leave in one click. The list is never rented, sold or shared, and 11,400 people have found that worth their inbox.
Put your own site in the box above and it joins the queue. If it gets picked, you read the findings the same morning everyone else does — no preview, no right of reply. That is the deal, and it is the only reason the teardowns are worth reading.
Every term below gets used loosely somewhere on the internet. These are the meanings I use in reports, so there is no ambiguity about what a number refers to. Expanded, with how each is measured →
Answer first, reasoning second — the same format I write client pages in, for the same reason.
GEO is the practice of getting a brand named inside AI-generated answers, such as those from ChatGPT, Perplexity, Google AI Overviews and Copilot. It works by making a clear, corroborated, machine-parsable claim about what a business is and what it is best at, then making sure that claim exists on the pages those systems retrieve from. It is sometimes called answer engine optimization (AEO) or AI search optimization, and the three names describe the same job.
Search engine optimization competes for a slot in a list of ten results; GEO competes to be one of two or three sources named in a single paragraph. The practical difference is where the work goes: GEO weights entity consistency, answer-first page structure, structured data and third-party corroboration far more heavily than link volume or keyword density. Semrush found around 90% of ChatGPT search citations go to pages ranking 21 or lower, which is why page-one rankings alone do not carry over.
First movement usually shows in six to ten weeks, and a meaningful change in citation share takes a full quarter. The fastest wins come from fixing entity inconsistency — one name, one category, one claim across your site, schema and third-party profiles — because models hedge or miscategorise you while sources disagree. Content and corroboration compound more slowly, which is why retainers here run six months minimum.
A fixed-scope visibility audit is $4,500 and takes three weeks. The ongoing Answer Engine Program is $9,800 a month with a six-month minimum, and the email-focused Owned Channel Build is $6,400 a month with a three-month minimum. Pricing is a starting point — scope sets the final number, and no retainer starts without a revenue baseline so both sides can tell whether the work moved anything.
With a prompt set, not a keyword list. We define 150 to 400 real buyer questions, weight each by the revenue of the customers who ask it, then run them against ChatGPT, Perplexity, Gemini and Google AI Overviews on a schedule and record whether the brand is cited and in what position. That citation data is joined to the CRM so the report shows revenue, not mentions.
No, and nobody honestly can. Retrieval indexes are not under any practitioner's control. What I guarantee is the inputs — the prompt set, the entity fixes, the extractable pages, the corroboration outreach and the email sequences — and honest monthly reporting on the outputs, including the months where a number moves the wrong way.
Yes, and they are cheaper run together. Technical health, crawlability and content structure are shared inputs: a page an assistant cannot parse is usually a page a crawler handles badly too. Ahrefs measured a 34.5% clickthrough drop for top-ranking pages once an AI Overview appears, so classic rankings still matter — they just no longer finish the job on their own.
You do not submit to it — you become retrievable. Three things decide it: your pages must be crawlable by the AI user agents (GPTBot, OAI-SearchBot, PerplexityBot and the rest, which many sites block by accident through a CDN rule), your claims must be stated in an extractable form rather than buried in narrative, and independent sources must corroborate them. In practice the blocked-crawler check takes ten minutes and resolves more cases than anything else I do.
Google AI Overviews first, because volume is not close — they appeared on roughly a quarter of all searches at 2025’s peak. ChatGPT second, because its users skew commercial and it drives the highest-intent referrals. Perplexity third for B2B and research-heavy categories. Copilot matters if you sell to enterprises on Microsoft 365. All four are tracked on the scoreboard regardless, because a brand that is retrievable for one is usually retrievable for the others.
It works particularly well, because local answers depend more on entity data than on content. Assistants building a local shortlist lean on your name, address, category and hours as they appear across your site, your schema and the directories — and those disagree at almost every multi-location business I audit. Reconciling them is unglamorous and usually the single highest-yield fix on the list.
Add it, but expect nothing from it on its own. It is a plain-text index of your best pages at the root of your domain, it takes an hour, and support across the major assistants is inconsistent. It is not a ranking factor and it does not compensate for an unclear entity or unextractable pages. This site has one; it is the least important thing on it.
Model roughly a third off your top-ranking informational pages where an AI Overview appears — Ahrefs measured a 34.5% clickthrough drop. Commercial and transactional queries are hit less, and AI Overview prevalence in your category matters more than the average. The compensation is that what still arrives converts far better: AI-referred visitors are worth about 4.4× an ordinary organic visitor, so the traffic chart falls while revenue per session rises.
A fixed-scope visibility audit is $4,500 and takes three weeks; the ongoing Answer Engine Program is $9,800 a month on a six-month minimum. Almost nobody publishes rates for this work, because almost every guide to it is written by a company selling software rather than the service. As a sanity check: below roughly $3,000 a month the daily tracking across four engines does not fit in the budget, so what you are buying is content volume with a new label on it. Full breakdown in the pricing table.
Because citation and recommendation are different outcomes, and only one of them is worth money. Ahrefs read 9,886 AI answers across four engines in 2026 and found that 43% of the answers citing their own page never mentioned their brand at all — the model used the source and named someone else out of it. It usually means your page states the category clearly and your differentiator vaguely, so a retriever lifting one passage gets the category from you and the recommendation from whoever else is named nearby. The fix is to put the differentiator in the same paragraph as the category term, and to get third parties saying it too.
It is the only responsible order, and it is what I insist on. Without a baseline nobody can tell month six from month one, and every claim about improvement becomes unfalsifiable. The audit produces the prompt set, the current citation share per engine, the entity conflicts, a technical read and a ranked fix list with a revenue estimate per item — which is enough to do the highest-value work yourself if you want to. It is credited against the first month if you continue.
Daily, and this is not a preference. In the Ahrefs experiment a cited page appeared on only about one day in three between its first and last appearance, so a single spot-check has roughly a one-in-three chance of describing your actual position. Anyone reporting AI visibility from a monthly manual check is reporting noise. The scoreboard here runs the full prompt set on a schedule and reports the trend, not a reading.
The entity work, yes — and you should start today whether or not we ever speak. Pick one name, one category and one claim, then make your site, your schema and your top twenty third-party profiles agree. What is genuinely hard to run in-house is the measurement: a weighted prompt set scored across four engines on a schedule and joined to your CRM is a build, not a spreadsheet, and without it you cannot tell a real gain from a reroll of the same question.
I check your visibility on twenty prompts from your category and send you what I find, with the revenue behind each one. You leave with the fix list whether or not you hire me.
These are modelled scenarios rather than signed client reports — each is calibrated against the published research on the overview, and each names the benchmark it is built on. Open one and drag the timeline.
Six builds: three we ship for clients, three we run ourselves and open up as they stabilise. Every one started as something a spreadsheet was doing badly for the fourth time. Tap any of them for the spec, the stack and who works on it.
No retainer starts without a prompt set and a revenue baseline — otherwise neither of us can tell whether the work moved anything. Pricing is a starting point; scope sets the number. What to expect at every budget →
| Visibility Audit | Answer Engine Program | Owned Channel Build | |
|---|---|---|---|
| Price | $4,500 one-off | $9,800/month | $6,400/month |
| Commitment | 3 weeks, fixed scope | 6 months minimum | 3 months minimum |
| Prompt set | 150–400, revenue-weighted | Same, re-weighted quarterly | — |
| Engines scored | ChatGPT, Perplexity, Gemini, AI Overviews | All four, daily | — |
| Cited and recommended tracked | Baseline only | Both, per prompt, joined to CRM | — |
| Entity and schema work | Audit and fix list | Built and shipped | — |
| Answer-first page rewrites | — | Ongoing | — |
| Corroboration outreach | Target list | Ongoing | — |
| Technical and content SEO | Technical read | Runs in parallel | — |
| Lifecycle and outbound email | Teardown | Included from month 6 | The whole engagement |
| Reporting | One written readout | Monthly, in revenue | Monthly, in revenue |
| Best if | You need to know where you stand before committing | You are losing shortlist decisions to an assistant | Traffic is fine; conversion and follow-up are not |
Why the prices are on the page. Every major guide to this discipline is published by a company selling software, so none of them will tell you what the service costs. That leaves the most common question in the category unanswered: what should I expect to pay for GEO? A fixed-scope audit runs $4,500 and takes three weeks. An ongoing program runs $9,800 a month on a six-month minimum. If you are quoted less than about $3,000 a month for “AI search optimization,” the tracking alone does not fit in the budget, and you will be paying for content volume with a new label on it.
What sets the final number. Category breadth, mostly — a prompt set of 150 questions is a different job from 400. After that: how many third-party profiles disagree about what you are, how much of your site can be edited without an engineering queue, and whether email is already instrumented. The audit exists so both of us can price the retainer against something real. It is credited against the first month if you continue, which is the only discount on this page.
$4,500fixed
$9,800/ month
$6,400/ month
Same shape on every engagement, so you always know what week you are in and what is due.
Analytics, CRM, email platform and search console. Nothing is written until there is a revenue number to beat and a prompt set to score against.
The AI pass over your customer data, the entity audit, the technical crawl and the email teardown. You get the ranked fix list with a revenue estimate per item.
Entity consistency, schema, crawl waste and the answer-first rewrite of your highest-intent pages. These are the changes that make everything after them compound.
Extractable answer pages and comparison assets go live while outreach starts landing you in the roundups and review platforms models retrieve from.
Welcome, abandonment, replenishment and win-back sequences, or the outbound rebuild — whichever your teardown said was leaking more.
Citation share by prompt, rankings, sessions and revenue, side by side with the baseline. Then we decide what the next quarter is for.
Saves us both a call.
Nothing compounding happens in 30 days. If the commitment is shorter than the feedback loop, the work is theatre.
Forty thin posts a month is what made AI search necessary in the first place. The model reads your data; it does not write your pages.
Nobody controls a retrieval index. I guarantee the inputs and report the outputs honestly.
If I cannot see revenue, I cannot tell you whether the program paid for itself — and that is the only number that matters.
Six questions covering the things that actually determine whether an assistant names you. It loads with example answers so you can see how it scores — change them to yours.
0/ 100
—
—Scored against the weighting used in a paid audit. A real baseline measures your prompt set against live model output rather than asking you.
What a program has to earn to be worth running. Enter your numbers; the model ramps improvement over twelve months rather than assuming it lands on day one.
Baseline With program Breakeven
Conversion-rate uplift assumptions are anchored to Semrush's finding that AI search visitors convert at roughly 4.4× the rate of ordinary organic visitors, discounted heavily because AI referrals are still a small share of total traffic.
Only a few engagements run at a time. The more specific you are about the gap, the faster I can tell you whether I am the right person for it — or point you at who is.
The same questions, one at a time, in a chat. Answer what you know and skip what you do not — it lands in my inbox either way.
Between the three, someone is working roughly 18 hours a day on weekdays. A brief sent overnight in the US is usually read before you wake up.
Either a candid “not a fit, here is why” or three questions I need answered.
I check your visibility on twenty prompts from your category and send you what I find.
We go through the findings and what each one is worth per month.
Sent straight to my inbox over an encrypted connection. Nothing is stored on this server and nothing is shared.