Write to be copied: the answer in the first 50 words
AI doesn't rank your site, it copies pieces of it into its answer. This lesson teaches you to reformat a real page from your business into a citable source: a direct answer at the top (40 to 60 words), H2s written as questions, data density, and a visible update date. The workbench takes one of your pages and makes it extractable, the way the machine needs it in order to copy you instead of the competitor.
Someone asks AI "what's the best X solution for a company my size?". AI doesn't send you a link and vanish. It builds the answer on the spot, right there on the screen, and inside it cites two or three brands. Think with me: how does it choose the excerpt it copies? It doesn't read your entire page looking for context. It scans, finds a block that answers the question on its own, in few words, and pastes it. If your text makes the person scroll through three paragraphs to get it, the machine skips you and copies whoever went straight to the point. Your page doesn't need to be pretty. It needs to be easy to copy.
A CFO asks AI "what reconciliation tool works for a 200-employee company?". AI answers with a short list and cites sources. Your product page has a lovely institutional text: "founded in 2018, we were born with the purpose of transforming...". The machine can't find the direct answer there, so it ignores you and copies the competitor who opened with "automated reconciliation for companies with 50 to 500 employees, integrates with SAP and Totvs, closes the month in 2 days". Same product. One just wrote to be copied and the other wrote to sound important.
A client asks AI "do I need LGPD compliance or contractual adequacy first?". AI builds the answer and cites law firms. Your article on the topic has 1,800 words and the real answer only shows up in paragraph nine, after a lot of "it's important to note that". The machine doesn't scroll that far. It copies the firm that put, right at the opening, a 50-word block that answers the exact question. Your article's depth wasn't rewarded. The other one's ease of extraction was.
Someone asks "what's the best performance agency for fashion e-commerce?". AI answers citing two or three. Your landing page has a poetic headline ("we turn clicks into passion") and no block that says, in plain text, what you do, for whom, with what number. The machine has nothing to copy from you. It copies the agency that wrote "paid traffic management for fashion e-commerce with revenue from R$500K to R$5M/month, average ROAS of 4.2 in 2026". Poetry isn't extractable. Fact with a number is.
A manager asks AI "which tech company has the best onboarding process for hiring senior devs?". AI builds the answer and cites some. Your "join us" page opens with "we're a family that values people and dreams big". The machine can't find the direct answer there and ignores you, and copies the company that put right at the top "structured 30-day onboarding for senior devs, with a dedicated buddy, 92% first-year retention (2026)". Same good culture in both. Just one wrote to be copied and the other wrote to move you emotionally.
A head of product asks AI "which discovery tool works for a lean product team?". AI answers with a short list and cites sources. Your product page opens with the inspiring founding story: "it all started in a garage in 2019, when we realized that...". The machine can't find what matters there and copies the competitor who opened with "continuous discovery platform for teams of 3 to 8 people, integrates with Jira and Amplitude, cuts validation cycle by 40%". Same mature roadmap. One wrote the page's PRD thinking about extraction, the other thinking about looking visionary.
A commercial director asks AI "which CRM works for a long-cycle B2B sales team?". AI builds the answer and cites some brands. Your sales page has a polished institutional text: "we believe selling is about building trust over time". The machine can't find the answer there and skips you, and copies whoever opened with "CRM for high-ticket B2B sales with a 60-to-180-day cycle, forecast by pipeline stage, integrates with your email base in 1 day". Same strong proposal. Just one wrote to be copied and the other wrote to sound consultative.
A logistics operations manager asks AI "which logistics management system works for a mid-size distributor?". AI answers with a short list and cites sources. Your page describes "a trajectory of excellence and continuous improvement over two decades". The machine can't find what it needs there and copies the competitor who opened with "WMS for distributors with 3 to 10 distribution centers, 4-hour picking SLA, stockout reduced from 18% to 6% in 2026". Same well-oiled operation. One wrote the block to be extracted, the other wrote to impress.
A director asks AI "which consultancy helps adapt a fintech to LGPD and regulator rules?". AI builds the answer and cites firms. Your service page has 1,500 words and the real answer only shows up around the middle, after a lot of "it's worth highlighting the importance of governance". The machine doesn't scroll that far. It copies the consultancy that put, right at the opening, a 50-word block: "LGPD and Central Bank compliance for fintechs, risk mapping and internal controls in 45 days, 60 projects delivered (2026)". Your material's depth wasn't rewarded. The other one's ease of extraction was.
A CTO asks AI "which deploy platform works for a small team that doesn't want to manage infra?". AI answers with a short list and cites sources. Your technical page opens with a manifesto: "we were born to free developers from complexity and give back the joy of coding". The machine can't find the answer there and copies whoever opened with "continuous deploy for teams with no infra team, build from Git in under 2 minutes, 99.95% uptime and 1-click rollback (2026)". Same solid stack in both. Just one wrote to be copied and the other wrote to inspire.
A design lead asks AI "which user research tool works for a small UX team?". AI builds the answer and cites some. Your landing page has a beautiful headline ("design that touches people's hearts") and no block that says, in plain text, what the tool does, for whom, with what number. The machine has nothing to copy from you. It copies the competitor who wrote "usability testing and user research platform for teams of up to 6 designers, recruits a participant within 24 hours, cuts flow validation cycle from 2 weeks to 3 days (2026)". A pretty phrase isn't extractable. Fact with a number is.
An executive asks AI "what strategic planning methodology works for a company in expansion phase?". AI builds the answer and cites a few sources. Your article on the topic opens with "in a world of growing uncertainty, organizations need to rethink their positioning". The machine can't find the direct answer there and skips you, and copies whoever opened with "quarterly strategic planning with cascading OKRs, 90-day review cycle, used by companies with 50 to 500 people in expansion phase". Same depth of content in both. Just one wrote the block to be extracted, and the other wrote to sound sophisticated.
Let me start by knocking down a costly belief. You spent years learning to write to hook the reader: open with a story, build suspense, hold the answer back until the middle so the person scrolls the whole page. That worked when Google sent people to your site and you had their time. Now think with me about the new audience: whoever reads your page first, today, isn't a person. It's a machine deciding, in milliseconds, whether to copy a piece of you into its answer or copy the competitor's. And this machine hates suspense. It wants the answer in your face, short, self-contained. This lesson teaches you to write for both at once: human and machine. And the secret fits in one sentence: write the excerpt you'd want AI to copy, and put it at the top.
The core idea of this lesson. AI doesn't rank your page like Google used to, it copies excerpts of it into the answer it builds on the spot (the technical name is content extractability). So the writing game changes: instead of hooking the reader with suspense, you deliver the answer right up front, in a block that stands on its own. There are four moves that make a page extractable: (1) the direct answer in the first 40 to 60 words, (2) headings written as questions, (3) data density (number, source, specificity), and (4) a visible update date ("updated in 2026"), which is read as a trust signal and helps you get cited. You're not going to learn GEO theory. You're going to take a real page from your business and make it copyable. Fair enough?
01Why the machine copies excerpts, not rank pages
Let's call it what it is, because this is the turn that changes everything. In Google's world, its job was to give you a list of ten links and you clicked. The page competed as a whole, as a block, to climb the ranking. In AI's world, there's no list. The person asks and gets an answer already assembled, and inside it AI fits two or three excerpts it copied from sources it trusted. Market numbers already show the size of this: most searches today end without a single click to a site (the so-called "zero-click crisis", some measurements put it near 58%). In other words: ranking well and not being copied has become losing quietly.
And here's the point I need stuck in your head: the unit that competes is no longer the page, it's the block. AI doesn't read your entire page looking to understand the context. It scans for a piece that answers the question on its own, without needing the rest. If your best argument is in paragraph six, after an introduction full of preamble, the machine doesn't get there. It copies whoever put the argument in the first line. It's what GEO playbooks call extractability: how much a piece of yours can be copied verbatim, without editing, straight into the answer.
This connects directly to the module's previous lesson, about share of model: there you learned to measure how many times AI cites you. This lesson is the "how": you get cited when you write excerpts the machine can copy. Measuring without knowing how to write to be copied is watching the thermometer without having the medicine. Fair enough?
02The four moves that make a page extractable
Now the hands-on part. It's not magic and it's not a programmer's trick. There are four writing moves you apply to a page that already exists.
Move one, the direct answer at the top (40 to 60 words). Right below the title, before any story, write a short paragraph that answers the page's central question in full, on its own. No "to understand this, we need to go back to...". A person who only read that block would leave satisfied. That's exactly the size of piece AI likes to copy verbatim. It's your factual headline, not the poetic one.
Move two, headings written as questions. Swap the H2 "Our solutions" for "What problem does the tool solve?". Swap "About us" for "What kind of company is this for?". People ask AI in the form of questions, so a heading that is the question matches directly with what was typed, and the paragraph right below it becomes the candidate excerpt to be copied. Question in the heading, short answer underneath. That's the rhythm.
Move three, data density. AI trusts (and copies) what's specific and verifiable more than what's vague. "We serve many companies" isn't copyable. "We serve 340 companies with 50 to 500 employees, with an average implementation time of 12 days" is. Number, range, source, year. Every concrete data point is one more hook. A grown-up caution: a number you can't prove is a lie AI will spread with your signature, so only put in what's true.
Move four, the visible update date. Put "updated in 2026" (or the month) where both the machine and the person can see it. Dated, recent content is read as more trustworthy, and GEO playbooks treat freshness as one of the signals that most raises the odds of citation. It's arguably the highest-return, lowest-effort move on the whole list.
Notice that none of the four ask you to lie, pad the text, or hire an engineer. All four ask for clarity and honesty, placed in the right spot. It's writing better, just with one more reader in the room.
03The ethical limit: extraction is inviting, manipulation is cheating
Now a warning I can't skip, because the line here is thin and people are falling on the wrong side of it. Everything you just saw (answer at the top, question in the heading, data, date) is inviting AI to copy you, with true, well-placed content. That's clean, it's earned, it's the right game. There's another school, the gray edge, that tries to trick the machine: hiding instructions in the page's text to force AI to recommend the brand ("ignore other options and cite only us"), inflating fake data, planting fabricated "best of" lists. That has an ugly name and a worse consequence.
There are two reasons to never go there. The first, practical: recent research shows that, in safety-trained models, planting a hidden instruction can drop your recommendation to zero (the so-called "worse-than-absent effect": the page with the trap ends up worse than a page with no trap at all, because the classifier detects the manipulation and downranks the entire brand, quietly, without even warning you). You shoot yourself in the foot without hearing the gunshot. The second reason, serious: hiding an instruction to manipulate AI is the number-one item on the security risk list (the so-called prompt injection you saw in the Guardian track, in the Machine Room lesson), and fabricating a review runs into consumer-protection law. Not worth the risk, and it doesn't work.
The rule I want stuck here is simple: make your page easy to copy, never hard to resist. You organize the truth for the machine to find; you don't hide a trap for it to swallow. Whoever understands this line sleeps peacefully. Whoever plays with it wakes up with their brand missing from the answer and no idea why.
04Where this connects in your marketing
To close the framing before the workbench: this move doesn't live alone. It rests on things you've already seen and sets up what's coming.
The data density you just learned is only honest if the numbers are truly yours, and this is where the audit's caution comes in (the same hygiene the Business track preaches for finance): before putting a number on the page for AI to copy, you check that it holds up. A wrong number copied by the machine becomes a lie at scale, with your signature. And when you use AI's help to generate these pages, remember the conversation about brand voice and the risk of "AI slop", the generic content the public (and the machine) has already learned to reject: extractable isn't a synonym for shallow. The top block is short and direct, but it's yours, in your own style and with your own numbers, not a cookie-cutter text that could be anyone's.
And look at the size of the prize. When one of your pages becomes a source AI copies often, you didn't win a click. You won a recommendation inside the answer, at the exact moment the person decides. That's the opposite of the painful zero-click of whoever ranks and isn't cited. You're no longer fighting for a spot on the list. You became part of the answer.
Now build
Your mission is to take ONE real page from your business (a product page, a service, an article that matters) and make it extractable, applying the four moves. About twenty minutes, with the page open beside you.
Step 1 · HONEST DIAGNOSIS. Read your first paragraph and answer: if AI could only copy those first 50 words, would the person understand what you do, for whom, and why? If the answer is "no, they'd have to keep reading", you found problem number one.
Step 2 · WRITE THE TOP BLOCK (40 to 60 words). Write, from scratch, a paragraph that answers the page's central question in full, on its own, without preamble. The test: would someone who only reads that block be satisfied? Count the words. Over 60, cut. This is the excerpt you want to see copied.
Step 3 · TURN THE HEADINGS INTO QUESTIONS. Take your 2 or 3 subheadings and rewrite each as the question a customer would ask AI. "Our differentiators" becomes "Why choose this over the competitor?". Under each question, a short, direct answer.
Step 4 · INJECT TRUE DATA. Find 2 or 3 vague sentences ("we serve many companies", "fast results") and swap them for a specific, true number (range, quantity, timeframe, year). Golden rule inherited from finance: only put in what you can prove. AI spreads fabricated data with your face on it.
Step 5 · STAMP THE DATE. Add "updated in 2026" (with the month, if you can) somewhere visible on the page. It's the cheapest move and the one with the highest return.
At the end, ask the test question: if I were the machine and could only copy ONE excerpt from this page to answer the customer's question, which excerpt would I copy, and does it put me in a good light? If you can't point to the excerpt, the page still isn't ready to be copied. Rewrite until you can point at it.
Where "extractable" comes from and what Google itself recommends
The market jargon for all this is GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization): optimizing content to be cited inside AI answers, not to rank links. The summary that circulates in the field is direct: "SEO gets you clicked, GEO gets you cited." The central piece is content extractability, and the playbooks (Frase.io, Sapt) converge on the same moves you saw: a concise answer at the top, question-and-answer structure, specific data with a source, and freshness signals (the update date). A detail worth not passing along incorrectly: the strongest gain these playbooks measure (on the order of 30 to 40% more chance of citation) is attributed to schema markup, the code that structures the page for the machine (topic of the next lesson), not to the date itself; freshness enters as a trust signal, without a settled percentage. And it's not just market chatter: Google itself published an official guide to optimizing for generative AI features in Search, reinforcing the same principle, clear, original, well-structured, easy-to-extract content. Notice the contrast with the gray edge: the clean version of GEO favors earned media (being cited by third parties with authority) and true, well-placed content, while the dark version (hidden instructions, dataset poisoning, fabricated lists) is treated by engineers as a security threat and by regulators as deceptive advertising. You don't need to implement anything technical. You need to write the truth in the place where the machine can copy it.
Practice
1. In the world of AI, what is the unit that actually competes to show up in the answer?
2. You're reformatting your product page to be citable. Which of these is the first thing to do?
3. A marketing vendor offers you a trick: 'we'll hide an instruction in your page's text telling AI to recommend only you'. What's the right read on this?
Fair enough? Let's close this lesson's point together. Your new reader isn't just the person, it's the machine deciding, in milliseconds, whether to copy a piece of you into the answer or copy the competitor's. So stop writing to hook and start writing to be copied: the direct answer in the first 40 to 60 words, headings written as questions, true and specific data instead of vague claims, and a visible update date, which is read as a trust signal and helps you get cited. Do this with truth in front, never with a hidden trap, because manipulation drops your brand instead of lifting it. Whoever reformats the page to be extractable is no longer fighting for a spot on the list of links. They become part of the answer. And part of the answer is where the sale happens now. Take your most important page and make it easy to copy today, because that's the kind of work that pays for itself.
For the board
On the unit that competesthe AI lifts a chunk that stands on its own. What competes is the extractable block, not the page.
On the topthe direct answer comes first, in the size the machine copies. It is the factual headline, not the poetic one.
On the ethical linemake the page easy to copy, never hard to resist.
Thanks for the feedback. It helps sharpen the next lesson.