Marketing after the click

The 97/6 problem

Ninety-seven percent of marketing leaders now use AI in their daily creative work, by Canva's count. Six percent have built it into how their team actually operates, by Supermetrics'. The distance between those two numbers is the entire story of marketing in 2027.

Access stopped being an advantage sometime in 2024, when every intern and every CMO got the same models on the same day for the same twenty dollars a month. What you compete on now is judgment: which problems you point the machine at, which of its forty ideas you throw away and whether you can tell finished work from work that's merely done. The tools are table stakes. The taste is scarce.

This guide is written for both ends of the org chart, the junior copywriter who suspects her job is changing shape and the CMO who has to explain to a board why the content budget should survive another year. It draws on current research, linked throughout so you can check my math, and on a dozen years of writing for brands while the ground moved underneath the industry twice.

The stakes are budgetary before they're existential. Nearly half of agencies expect cuts this year, by Onclusive's survey of the PR side, and the teams that survive reviews are the ones who can draw a straight line from their work to revenue. AI won't draw that line for you. Used well, it will hand you the pen.

A note on what you won't find below: tool recommendations. Any list of products would be stale by the time you finished reading it, and the teams that win aren't distinguished by their software anyway. Everything in this guide is a practice, which means it survives the next model release, and the one after that.

The scoreboard changed

For twenty years the deal with Google was simple. You made pages, Google sent people, everyone counted clicks. That deal has ended, and the numbers describing its end are blunt. When an AI Overview sits on top of the results, people click a traditional link about 8 percent of the time, down from 15 percent without one, by Pew's tracking of nearly 69,000 real searches. Seer Interactive watched organic clickthrough on those queries fall 61 percent across 25 million impressions. Similarweb had zero-click searches at 69 percent of all Google queries by mid-2025, and Bain's consumer research puts the no-click rate at 83 percent whenever an AI answer appears.

Read those numbers carelessly and you'll cut the wrong things. What collapsed was the casual click. The person who once skimmed your blog post to settle a passing question now gets settled by the machine, while the person who still clicks through arrives pre-sold. Semrush's analysis of 500 high-value topics found that visitors referred by AI tools convert at 4.4 times the rate of traditional organic visitors. Seer measured ChatGPT referrals converting at 15.9 percent against 1.76 percent for Google organic. AI referral traffic remains a sliver, around 1 percent of sessions across the 13,770 domains Conductor tracks, but it grew more than 500 percent year over year and it's the most qualified traffic most websites have ever received.

The strangest number comes from BrightEdge. In late 2024, roughly three quarters of the sources cited in Google's AI answers also ranked in the organic top ten. By early 2026 that overlap had collapsed to somewhere between 17 and 38 percent, which means five of every six citations now come from pages that rank nowhere near the top. Ranking and being cited have divorced, and they're not sharing custody of your strategy. You need a plan for each.

Getting cited in 2027

Search optimization and AI optimization are one discipline now, viewed through two lenses, and every credible forecast says the lenses keep merging as search and assistants blur into a single behavior. Keep the proportions in mind before you rebuild anything. Google still handles roughly nine in ten searches, classic rankings still pay the bills and 99.2 percent of the queries that trigger AI Overviews are informational, by Ahrefs' keyword data, so your commercial pages still live and die by ordinary SEO. Everything below assumes you keep doing the basics well. The additions are about becoming the source machines quote.

3.1Build the entity before the keyword

AI engines cite between two and seven domains per answer, by Search Engine Land's analysis. Two to seven. That's the entire shelf, and shelf space goes to entities the machine recognizes and trusts. Recognition is built the unglamorous way. Keep your company name, founding facts and expertise claims consistent across your site, your LinkedIn, your directory listings and every third-party mention. Schema markup, Organization and Article at minimum, makes those facts machine-readable instead of machine-guessable. Author pages need real credentials attached to real names. A brand the model can resolve gets cited. An ambiguous one gets paraphrased without credit, which is the 2027 version of ranking on page nine.

3.2Treat PR as a ranking discipline

The engines favor earned, third-party sources over anything you publish about yourself, which turns digital PR from a brand exercise into a direct visibility lever. SE Ranking's research found that domains with living review profiles on G2, Capterra and Trustpilot are three times more likely to be cited, and heavily referenced sites are 3.5 times more likely than obscure ones. Your org chart probably keeps PR and SEO on separate floors. The machines merged them years ago, and your budget should catch up before your competitor's does.

3.3Write for the passage economy

A machine assembling an answer never grades your page. It grades the 120 words it lifted. That changes composition. Open each section with the direct answer to the question its heading poses, then earn the depth underneath. Heading hierarchies stay clean so retrieval knows what each passage claims to be. Question-format headings and honest FAQ blocks earn their keep where the material suits them, and a date stamp on every substantive update feeds the freshness signals the engines watch. None of this requires writing for robots. A well-edited explainer already works this way, which is the joke at the center of the whole discipline. The machines reward what good editors demanded all along.

3.4Publish what only you know

Language models need sources for claims, and the surest way to be the source is to originate the claim. Run the annual survey. Publish the pricing benchmark. Release the dataset from your own operations with your name on the methodology. One proprietary study will earn more citations than fifty explainer posts, because the explainer has ten thousand substitutes and the benchmark has none. This is also the cheapest moat left in content, since it requires effort rather than budget, and effort is the one input your competitors' AI subscription doesn't include.

3.5Keep the plumbing boring and open

Fast pages, clean architecture, mobile discipline and schema markup remain the entry fee. Add one check most teams skip: open your robots.txt and confirm you're admitting GPTBot, ClaudeBot and PerplexityBot. Plenty of companies pay for AI-visibility dashboards while their own config turns the crawlers away at the door, a problem no dashboard will ever report because the dashboard can't see a site the crawler never fetched.

3.6File llms.txt under afternoon projects

The proposed standard where you hand AI systems a curated map of your site had a moment, and the evidence since has been unkind. Ahrefs crawled 137,000 domains and found 97 percent of llms.txt files received zero requests. Google declined to support it. OpenAI points people to robots.txt instead. Perplexity claims to read it, and coding assistants do, which matters if you sell to developers and barely otherwise. Publish one if you like tidy websites, keep it current or delete it, and expect nothing from it before 2028.

3.7Get ready for shoppers who aren't people

The next shift is already visible in the research firms' scenario decks: assistants that check availability, compare options and complete purchases on a person's behalf. Harvard Business Review sketches three modes of it, from brand-built agents to personal shopping agents to full agent-to-agent commerce. What an agent needs from you is structured, accurate, real-time data: price, stock, hours, booking, specs. What it will punish is drift between your marketing claims and your operational record, because agents aggregate reviews and complaints at a scale no human researcher ever managed. BCG already counts 43 percent of consumer journeys as research-led, with AI tools ranked among the most influential touchpoints by the people who use them. The brands that win agent recommendations will be the ones whose promises survive an audit, which is a strange and healthy thing for marketing to be graded on.

3.8Show up where the engines watch video

The camera and the microphone are becoming query boxes. Credible forecasts for 2027 put voice, image and video search on the mainstream path, with someone pointing a phone at a failed water heater and asking who can fix it locally treated as an ordinary Tuesday behavior. Assistants already lean on video platforms and transcripts when they assemble answers, which turns work you made for engagement into source material for engines that can finally watch it. The practical moves are small. Publish transcripts alongside every video, since the transcript is what gets read. Give product and how-to images descriptive filenames, alt text and markup. Keep location, hours and service data structured so the water-heater query resolves to your business instead of a directory page scraping your business.

3.9Measure citations the way you measured rankings

Build a citation share-of-voice habit: a recurring check of what ChatGPT, Gemini, Claude and Perplexity say when asked the questions your buyers ask, tracked over time like a rank report. Segment AI referrals into their own analytics cohort, since averaging 15.9 percent conversion traffic into 1.76 percent traffic hides your best channel inside your biggest one. The measurement tooling matures fast through 2027, and there's a bonus for showing up: Seer found brands cited in AI Overviews earn 35 percent more organic clicks on the very same results page. Presence in the answer reads as endorsement, even to the people who scroll past it.

The sharpest version of this practice is answer-gap analysis. List the fifty questions that precede a purchase in your category, ask each assistant all fifty and note where competitors get cited while you don't appear. Every gap is a brief: a passage to write, a review profile to build, a piece of earned coverage to pursue. Rank the gaps by deal value rather than query volume, since a citation on a question asked eighty times a month by the right buyers beats one asked eighty thousand times by students writing papers.

The median, and how to leave it

The standard complaint about AI creativity, that it regresses to the mean, is accurate and also the most useful thing about it. A language model is a map of what everyone else would write. Owning that map is worth a great deal to anyone whose job is to write something else.

4.1Map the cliché, then leave it

Before writing a single headline, ask the model for forty. The first ten will be what every competitor's AI wrote this morning. The next twenty will be variations on those ten. Somewhere past thirty the model starts to strain, and that strain marks the border of the obvious. Your brief becomes everything the list didn't contain. A census of the predictable used to take a decade of pattern exposure to carry in your head. It now takes ninety seconds, and most marketers use those ninety seconds to pick option three.

4.2Make it argue against you

Models are better editors than authors, because critique runs on pattern recognition, which they have in oceans, rather than on stakes, which they lack entirely. Hand over your draft and ask it to kill the idea. Ask it to impersonate your most skeptical client, your CFO, the journalist who'd mock the press release. Ask which claim in the piece is weakest and which paragraph a reader would abandon. You'll reject half the critique, and the half you keep would have cost you a bruising meeting to learn any other way.

4.3Interview before you draft

The largest quality gap in AI-assisted writing sits upstream of the writing. A model drafting from a topic produces the average of the internet. A model drafting from a recorded interview, with real names, real numbers, a position on a live debate and a story that cost the teller something, produces a piece no other company could publish. I've built my practice on this sequence, it's the core of the De-AI process I run every client piece through, and it remains the least copied move in the industry because it looks slow. Measured against three rounds of revisions on a draft nobody believes in, it's the shortcut.

4.4Test on synthetic audiences, decide with real ones

AI personas built from real customer data can now pressure-test concepts before a human ever sees them. One 2025 experiment found digital-twin respondents matching real survey results with 94 percent accuracy, and the honest read of the field is that synthetic panels are excellent at flagging weak ideas and untrustworthy at anointing strong ones. The models skew young, educated and agreeable, and they're confident when they're wrong. Use them as a filter that saves your real research budget for the concepts that survive.

4.5Build voice from transcripts, never from adjectives

Telling a model to write "confident but warm, professional yet approachable" produces the same beige it produces for everyone who types those words, which is nearly everyone. A voice lives in specifics: sentence length habits, pet phrases, what gets a joke and what doesn't. Twenty minutes of someone's transcribed speech, annotated with the tics worth keeping, outperforms any adjective list ever written. For executive ghostwriting this is now the entire job. Record the interview. The transcript is the style guide.

4.6Keep a ledger of your rejections

Every time you kill a machine draft, write down why. Too smooth. Wrong register for this client. The claim in paragraph two would never survive legal. Within a quarter that ledger becomes three things at once: a custom instruction set that makes your next hundred drafts start closer to right, a training document for every new hire who joins the team and a written record of your editorial judgment, which is the asset your whole price is built on. Most teams let this knowledge evaporate at the end of each chat session. The ones who capture it are compounding while everyone else starts from zero every morning.

4.7Give it strange constraints

A model's obedience is a creative asset the moment the instruction gets interesting. Ban its twenty favorite words and watch the prose change texture. Demand the piece read like a field manual, a menu, a deposition. Force the analogies to come from aviation or botany or nineteenth-century shipping. Constraint has always been where creative work gets its shape. The difference now is that testing a constraint costs a minute instead of a morning, so you can afford to try ten of them before lunch.

Where the money hides

Supermetrics found 87 percent of marketers using AI for content creation and 39 percent using it for reporting and analytics. Hold that pairing up to the light, because it's upside down. Content is where machine output faces its most hostile audience, consumers who insist they can smell the difference. Analysis is where machine output compounds without a single person objecting. Nobody has ever bounced off a quarterly report because the pivot tables felt soulless.

The unglamorous applications carry the return. Transcribe every client and sales call, then treat the archive as a searchable database when renewal season arrives. Feed a quarter's media coverage into a model and ask what changed. Half of marketing teams wait one to three days for answers to basic data questions, by the same Supermetrics report. A marketer with a model wired into the numbers waits about a minute, and the compounding value of a thousand faster decisions beats the value of a thousand faster blog posts by an embarrassing margin.

None of it works on a broken foundation, and the report's phrasing deserves its small fame: you can't outrun a bad data foundation. The 6 percent who embedded AI properly share habits worth stealing. They define the use case before buying the tool, they keep data strategy inside marketing instead of ceding it to IT and they treat metrics as signals rather than gospel. Meanwhile 37 percent of marketers report no clear AI strategy from leadership at all, which reads as a complaint and is in fact a vacancy. Write the one-page strategy nobody else has written. Promotions have a way of finding the person who ends an ambiguity.

For agencies and freelancers the business-model question is louder than any workflow question. When drafting costs collapse toward the price of electricity, billing by the word or the hour advertises that you sell the part a machine now does. The market is repricing writing in two directions at once. Commodity output is heading to zero while accountable judgment, the interviews, the strategy, the voice work, the willingness to sign your name to a claim, keeps getting more expensive. Price against the second market. Retainers built around editorial governance, a defined voice system and a person who answers for the work will survive procurement conversations that per-word pricing cannot.

There's a staffing riddle attached, and it lands on both readers of this guide at once. AI absorbed the tasks juniors used to learn on, the first drafts and the research memos, which raises an uncomfortable question about where the senior editors of 2032 will come from. The teams solving it have flipped the apprenticeship. Juniors start in the critic's chair, running the claims passes, the citation audits and the model red-teaming, and they edit machine drafts against a senior's margin notes until the margin notes become instinct. Editing was always where writers actually got made. The industry just spent thirty years pretending drafting was the classroom.

Quality control is where AI-heavy teams separate. Mentions of "AI slop" grew ninefold in a year, by Canva's tracking, and 41 percent of marketing leaders concede the problem lives in their own house. The fix is systematic rather than aspirational: keep a banned-phrase list and enforce it, run a claims pass where every factual statement gets a source or gets cut, version your prompts like creative briefs with named owners. A team that treats prompts as tribal knowledge scattered across personal chat histories is running a kitchen where every cook keeps the recipes in their head. It works until the Tuesday it doesn't.

One habit belongs on the calendar above all of these. Once a quarter, ask the four major assistants what they'd tell a buyer about your brand, your pricing, your category and your competitors. Pernod Ricard ran that exercise and found leading models carrying incomplete or wrong information, including an affordable whiskey miscategorized as a prestige label. An error like that, repeated across millions of private answers, is a positioning problem no campaign can reach. The remediation happens upstream, in the sources models learn from: your structured data, your review profiles, the reference sites that describe you. Fix those, then re-audit next quarter. Almost nobody does this yet, which is precisely the argument for starting.

Ten things most marketers still don't know

Collected from the research and from watching smart teams get surprised by each one.

1. Google runs AI Overviews on 88 percent of healthcare queries and about 4 percent of e-commerce queries, by BrightEdge's tracking. It automated the library and kept the cash register human. Your AI-search exposure depends on your vertical more than on your strategy, so check your own category before copying anyone's playbook.

2. Consumers grade on feel. Seventy percent say they can spot AI-generated ads because something is missing, and 78 percent prefer human-made work even when the AI version might be technically superior, by Canva's consumer research. Production method is invisible. Emotional register isn't.

3. Roughly half of AI answers volunteer comparisons and recommendations nobody asked for, by BCG's analysis. A buyer researching your product will meet your competitor inside your own answer, sometimes with an endorsement attached. Your battlecards now need to work in absentia.

4. Assistant queries carry motive in a way keywords never did. People tell a model to plan the ten-day anniversary trip and mention they're pescatarian. If you run a site assistant or chat interface, its logs hold richer customer insight than your keyword reports ever will, and most teams never read them.

5. AI detectors are guesswork wearing a lab coat. OpenAI retired its own classifier over low accuracy, and Stanford researchers showed detectors disproportionately flagging non-native English writers. Grade writing on whether it's true, specific and worth a reader's time, because that's the only detector that has ever worked.

6. Long, question-shaped queries trigger AI answers far more often than short ones, by Ahrefs' analysis. The head terms you bid on are the least exposed to AI answers. The long-tail informational content you deprioritized years ago is the most exposed, and the most citable.

7. Gartner projected traditional search volume falling 25 percent in 2026 as usage migrates to AI platforms. Forecasts miss, and the direction hasn't wavered in three years. Shift your budget mix ahead of the trendline rather than behind it, because behind it is where the case studies about disrupted incumbents get written.

8. Ninety-six percent of CMOs say AI is transforming their function end to end. Thirty-one percent have implemented agentic execution, by BCG's count. That 65-point gap between narrative and operations is where consultants are currently getting rich, and where an operator who actually ships things gets noticed.

9. Only about 1 percent of users click the links inside an AI Overview, by Seer's measurement. The answer box is a branding surface rather than a traffic source. Write your citable passages knowing the mention itself is the payoff, the way a broadcast mention was.

10. ChatGPT alone accounts for 87 percent of AI-sourced referral traffic, by Conductor's data across 13,770 domains. Optimizing for "AI search" today mostly means optimizing for one product. Anyone who lived through two decades of Google dependence should recognize the shape of that risk immediately.

Comms in the agent era

Corporate communications enters 2027 squeezed from both sides. The press corps is thinning, with roughly half of PR practitioners expecting fewer journalists on their beats, by Onclusive's survey, while the journalists who remain drown in machine-written pitches that all read like each other because they are each other. The response is depth over volume. A working list of twenty tier-one relationships beats a database of two thousand contacts, and the pitch that wins carries something a model can't generate: a number the journalist can't get anywhere else, an executive willing to say the impolitic thing on record, access.

Distribution moved too. Around 65 percent of people now get news through social video, which makes an executive's face and voice a media channel whether the executive likes it or not. The ghostwritten LinkedIn essay still has a job, and it no longer works alone. Comms teams that can produce a credible talking-head minute, a podcast appearance and a keynote excerpt from one interview session hold a structural advantage over teams still routing everything through text.

Under both shifts sits the same asset. Commodity writing has collapsed to the price of electricity, while voice-matched, synthesis-heavy, position-taking communication keeps appreciating. An executive with an actual point of view, extracted through real interviews and developed across years of consistent public argument, is the one asset a competitor can't prompt into existence. Executive voice deserves a line in the budget and a defensible process behind it, because it's simultaneously the most valuable surface in your comms program and the easiest one to erode with a lazy quarter of generated posts.

The audience for all of this now includes machines, and the machines fact-check. Agents and assistants aggregate your reviews, your support threads, your employee commentary and your forum presence into the answers they give about you, which means the gap between what you promise and what customers experience is machine-legible and gets quoted back to buyers. BCG frames this as brand stewardship expanding to cover operational truth. A comms plan that stops at the press release now covers about a third of the record the machines are reading.

Crisis work inherits the same physics at higher speed. The old playbooks assumed a news cycle measured in hours and a correction that ran where the error ran. Answers now assemble in real time, and a false claim that reaches the models gets repeated inside millions of private conversations no monitoring dashboard can see. The counter is a correction path built in advance: an authoritative statement published fast at the places machines re-check, your newsroom page, your structured data, the reference sites that describe you, plus a daily re-run of the quarterly brand audit until the answers come back clean. Speed still matters in a crisis. Being machine-legible while you're fast is the new part.

Storytelling itself changes less than the trade press suggests. Google's quality framework put experience first among its signals for a reason, and every mechanism described in this guide points the same direction. Stories with named people, real dates, real numbers and consequences that cost somebody something work on human readers, earn machine citations and survive skeptical editors. Write the case study with the client's actual figures or don't write it. The middle path, the vague success story with a stock photo and a rounded percentage, now fails with every audience it has.

The existence test

One diagnostic from my own framework is worth handing over in full, because it grades everything else in this guide. The most reliable tell of machine writing is that nothing in the piece requires the author to exist. No detail that had to be lived. No position that costs anything. No number that came from inside the building. A model produces work like that by default, since there's no author in there and nothing was ever at stake. A human produces work like that on deadline, on autopilot, on the fourteenth blog post of the month.

The cure is the same in both cases. Put something in the work that could only have come from you. That's what the machines cite and what the buyers remember, and it's the entire test. It fits on a sticky note. It should probably go on yours.