Growtika
    LET'S TALK
    The human layer of AI visibility

    How to optimize your company for AI search: the human checklist

    Everyone is buying tools to win at AI search. The companies actually getting cited are doing the opposite: more human work, not less. Here is the checklist of that work.

    Yuval Halevi•July 16, 2026•16 min read
    AI answer anatomy
    Synthesized answer

    The brand becomes the answer only after the web agrees.

    Clear positioning, named expertise, independent mentions and original proof create the confidence a model needs to cite you.

    Source 01Source 02Source 03
    PublicationIndependent mentionA trusted source says the claim for you.
    ExpertNamed authorityA real person stands behind the point.
    ProofOriginal evidenceData and outcomes competitors cannot copy.
    CommunityReal consensusUseful participation, not manufactured noise.
    Human work
    creates
    machine trust

    6 things to remember

    • Tools optimize the surface. Authority, credibility, and consensus, the things that actually get you cited, are human work no dashboard does for you.
    • Decide what you want to be known for before you touch a single page. Ambiguity is the main reason AI describes you wrong.
    • Third-party mentions beat on-site optimization. AI trusts what others say about you more than what you say about yourself.
    • Real, named author expertise now shows up in citations. Anonymous "content" gets averaged into background noise.
    • Original data is the one thing a competitor cannot copy and a model cannot hallucinate around. Make numbers nobody else has.
    • Audit what the machines say about you monthly. Wrong answers spread through consensus, so fix the source, not the symptom.

    For the past year I've watched companies react to AI search the way they reacted to every channel before it. Buy a tool. Automate the thing. Set it and forget it. From what we've seen across client work, that instinct is backwards here. The layer these tools optimize is the layer that matters least.

    The parts that actually decide whether ChatGPT, Claude, and Perplexity recommend you are stubbornly, inconveniently human. Someone has to earn the mentions. Someone has to be the named expert. Someone has to run the research nobody else has run. You cannot schema-markup your way into being trusted.

    So this is not another list of technical tweaks. It's the checklist of human work that moves the needle, organized into six jobs a machine cannot do for you. Some of it is slow. All of it compounds.

    more AI in the loop means more human work at the base. that's the whole article.

    Why AI search rewards human work

    Traditional SEO trained us to think in terms of things you do to a page. Titles, headings, internal links, markup. AI search flips the model. A language model does not rank your page so much as assemble an answer from what the wider web appears to agree is true about you. That shift moves the center of gravity off your own site and onto everyone else's.

    Which is exactly why most "AI visibility" tools disappoint. They track and format. They tell you that you're invisible and dress your pages up nicely. What they cannot do is the underlying work: getting an independent publication to mention you, turning a real engineer into a quotable source, producing a number no competitor has. That work is done by people, and it's the part that actually earns a citation.

    How the citation is actually built
    01
    Independent authorityIndustry publications and trusted roundups
    02
    Named expertiseA real person with a visible track record
    03
    Original proofData, benchmarks and customer outcomes
    04
    Community consensusReal conversations across the wider web
    AI ANSWER

    “The strongest option for teams that need clear category expertise and evidence from independent sources.”

    [1][2][3]
    The machine assembles. People create the evidence.
    SchemaFormattingTrackingAlerts Useful, but only after the trust exists.

    Updated: July 2026

    The job What a tool can do What only a human can do
    Positioning Flag that your description is inconsistent across platforms Decide what you should actually be known for
    Third-party mentions Find the pages that already cite you Get you added to them, with a real reason to
    Author credibility Check whether author markup exists Be a named expert with a track record worth trusting
    Community Monitor where you're mentioned Show up and earn trust as a real participant
    Original data Format it and track its pickup Generate numbers nobody else has
    Wrong info Detect that the answer is wrong Trace and fix the source it came from

    None of this means tools are useless. A good one points your attention at the right gap and saves you hours of manual checking. It just cannot close the gap for you. The mistake is treating the dashboard as the work instead of the map.

    How most companies play it

    Buy an AI visibility tool. Add schema. Rewrite meta descriptions. Watch a score that will not move because the underlying trust was never built.

    Six months later: still invisible, now with a subscription.

    What actually works

    Decide your story. Earn mentions on pages AI already cites. Put real experts forward. Publish data only you have. Then let a tool tell you it's working.

    Slower to start, but it holds once it's built.

    Before you automate, ask four questions

    If you are evaluating a tool for this, these questions separate the ones worth paying for from the ones selling a number.

    Question to ask Why it matters What good looks like Red flag
    Does it build authority, or just measure it? Measurement is cheap. Trust is the hard part. Clear picture of what is actually missing "Guaranteed AI visibility"
    Can it show the wrong facts and where they came from? Fixing the source is the real work Traces citations back to origin pages Only shows a headline score
    Does it separate "are we visible" from "are our sources trusted"? Those are two different problems Two distinct signals you can act on One blended number that means nothing
    Would it replace human work, or point it? Nothing on this list is fully automatable Frees your people to do the earning Promises to do all of it for you

    The six human jobs

    Here is the whole checklist at a glance. Jump to any job, or use the picker below to find where to start based on what AI is doing to you right now.

    Not sure where to start?
    Pick what AI is doing to you right now
    Your first three moves
    Pick a situation on the left and I'll show the three things to do first, in order.
    Based on patterns from 50+ B2B companies
    1

    Decide what you stand for

    Ambiguity is the number one reason AI describes you wrong

    Before any of the outreach or research, someone has to make a decision that no tool can make for you: what should the machines say when a buyer asks what you do? If the answer is fuzzy inside your own company, it will be fuzzier in the model. Most "AI got us wrong" problems trace back to this.

    1 Get the story straight

    Write your one-sentence category claim in plain language
    The exact sentence you want AI to repeat when someone asks what you do. No adjectives you cannot defend. If three people on your team write it differently, that is your first problem to solve, not a keyword gap.
    High Impact Low Effort 1-2 hours
    Pick the 2-3 facts you want tied to your name
    Not ten. Two or three specific, defensible claims: who you serve, the one outcome you deliver, the one thing you do differently. These become the exact statements you build consensus around in Job 2 and Job 5.
    High Impact Low Effort 1 hour
    Cut the positioning you cannot back with proof
    Every vague superlative you cannot support is a claim a model will either ignore or contradict from a more credible source. "Leading platform" with nothing behind it is worse than a modest claim that checks out.
    Medium Impact Low Effort 1 hour
    2

    Earn the mentions

    AI trusts what others say about you more than what you say about yourself

    This is the job most companies skip and the one that matters most. A model assembles its answer from independent sources. One mention from your own site is marketing. The same claim echoed across a review site, a Reddit thread, and an industry publication starts to read as fact. That echo is built by people, one relationship at a time.

    2 Build independent sources

    Get onto the pages AI already cites for your category
    Run your buyer queries, note which third-party pages get cited (comparison posts, roundups, listicles), then get yourself added or updated on them. From what we've seen, infiltrating a page a model already trusts is far faster than ranking a new one.
    High Impact Medium Effort 2-4 weeks
    Turn one real expert into a quotable source
    HARO and Qwoted responses, podcast appearances, guest analysis. One named person answering journalists' questions creates independent mentions no tool can manufacture. Pick the person, protect their time, make it a habit.
    High Impact Medium Effort Ongoing
    Make partners and customers say your name in public
    Joint case studies, an integration page on their site, honest reviews. Each one is an independent source repeating the same claim. This is the least glamorous and most durable mention-building you can do.
    Medium Impact Medium Effort Ongoing
    Growth Hack

    Update, do not beg

    Do not email an author asking to be added to their roundup. Bring them something they want instead: a fresh data point, a factual correction, an exclusive stat for their piece. Editors update pages for value, not favors.

    The same outreach earns you a backlink and an AI citation from a page the model already trusts. One motion, two wins.

    3

    Put real people forward

    Anonymous content gets averaged into background noise

    Models are swimming in text. What they are increasingly short on is a signal that a real person with real expertise stands behind a claim. That signal is easy to add and most companies still do not. Faceless "content" is the easiest thing in the world to skip.

    3 Make expertise visible

    Attach a real, credentialed author to anything you publish
    Name, face, short bio, linked profiles. From what we've seen, the same article tied to a named expert with a visible track record gets pulled into answers more often than when it runs anonymously or under a brand byline.
    High Impact Low Effort Ongoing
    Write from first-hand experience, not a summary of the top ten results
    The one thing a model cannot generate is what you actually saw: the test you ran, the thing you shipped, the deal you lost and why. Rephrased consensus is exactly what the model already has. Your scar tissue is the differentiator.
    High Impact Medium Effort Ongoing
    Take a real, defensible position
    A clear opinion gets quoted. Hedged, everything-to-everyone copy gets averaged into the background. You do not need to be loud, just specific enough that a model has something concrete to attribute to you.
    Medium Impact Low Effort Ongoing
    From My Experience

    The pages of ours that get cited most are not the comprehensive ones. They are the ones where a named person says something specific and slightly uncomfortable that the rest of the internet keeps hedging on.

    Consensus is the floor. It gets you into consideration. A defensible, non-obvious take is what gets you pulled out of the crowd and quoted by name.

    4

    Join the real conversations

    The places your buyers actually talk are training data now

    A large share of what models know about your category came from communities: Reddit, niche forums, Slack and Discord groups, Q and A sites. You cannot fake your way into those. You can only show up as a real person who is useful. That is slow, human, and largely un-outsourceable, which is exactly why it works.

    4 Show up where it counts

    Be a genuine participant in the 2-3 communities your buyers use
    Not a poster dropping links, a person who answers questions and gets recognized. Pick the two or three that matter for your category and go deep. Breadth here reads as spam. Depth reads as authority.
    Medium Impact High Effort Ongoing
    Answer the questions your sales team keeps hearing, in public
    The questions from sales calls and support tickets rarely show up in keyword tools, but they are exactly what buyers type into AI. Write real answers where both humans and crawlers read them. These zero-volume questions convert far above their apparent size.
    High Impact Medium Effort Ongoing
    Never astroturf
    Fake reviews and sockpuppet threads are the fastest way to teach a model the wrong thing about you, and they are increasingly easy to detect. One real advocate is worth more than ten fake accounts, and carries none of the risk.
    Foundation Low Effort Ongoing
    Warning

    Astroturfing backfires twice. Platforms remove it, and if it does survive, you have seeded a coordinated pattern that reads as manipulation to the exact systems you are trying to influence.

    Consensus only works when the sources are genuinely independent. Manufacturing that independence is the one shortcut that actively sets you back.

    5

    Make proof only you have

    The one input a competitor cannot copy and a model cannot invent

    Everything else on this list can, in theory, be matched by a competitor with enough patience. Original data cannot. A number that exists nowhere else gets cited precisely because it exists nowhere else, and it pulls your name into answers about the whole category, not just about you. This is the highest-leverage human work here.

    5 Generate original proof

    Publish one piece of original data per quarter
    A survey of your market, a benchmark you ran, an analysis of your own anonymized product data. It does not need to be huge. It needs to be yours. Original numbers earn citations because no competitor can serve them up instead.
    High Impact High Effort 1-4 weeks per study
    Put real customer outcomes in public, with specifics
    Named results with real numbers beat "trusted by leading teams." Specific, verifiable claims are what a model is willing to repeat. Vague social proof is invisible to it. Get permission, then get specific.
    High Impact Medium Effort Ongoing
    Disclose your method and any conflicts up front
    Sourced, transparent data survives scrutiny and gets referenced. Unsourced numbers get ignored or, worse, flagged. State your sample, your method, and what you sell before anyone has to ask. It is also what keeps the research credible on Hacker News.
    Medium Impact Low Effort Built into each study
    The Bottom Line

    Original research is the single highest-leverage item on this list. It is the one input a better-funded competitor cannot simply outspend you to replicate, and the one thing a language model cannot invent when it builds an answer about your space. If you do only one hard thing this quarter, do this.

    6

    Audit the machines

    Wrong answers spread through consensus, so fix the source, not the symptom

    You cannot manage what you never look at. The most useful, most neglected habit in AI search is simply asking the models the questions your buyers ask, then reading the answers closely. This is your real ranking report now, and it takes about two hours a month.

    6 Run the loop

    Run 20-30 buyer queries across ChatGPT, Claude, and Perplexity every month
    The exact questions your buyers ask, not vanity searches for your brand name. Record how you are described, what is wrong, and who gets recommended instead. Keep it in one spreadsheet so you can see movement over time.
    High Impact Low Effort 2 hours per month
    When AI states something wrong, trace it to the source
    A model repeating a wrong fact learned it somewhere: an outdated page, a bad review, a stale third-party profile. Editing your homepage does nothing if the wrong claim lives on G2 or an old press release. Fix the origin.
    High Impact Medium Effort Varies
    Watch the citations, not just the answer
    Note which domains AI keeps pulling from for your category. Those are your Trust Hub, and they are field-specific: what AI trusts for cybersecurity is not what it trusts for developer tools. Getting onto those domains is the whole game.
    Medium Impact Low Effort 1 hour
    The queries to actually run

    Swap in your category, use case, and competitors. Run each one in ChatGPT, Claude, and Perplexity, then log the wording, the citations, and who gets named instead of you.

    1. best [category] for [specific use case]
    2. top [category] tools for [buyer type]
    3. is [your brand] a good fit for [use case]
    4. [your brand] vs [main competitor]
    5. alternatives to [competitor everyone knows]
    6. which [category] tool do people actually recommend for [pain point]
    The fix-the-source loop
    1Ask real buyer questionsRun the exact prompts prospects use
    →
    2Find the wrong claimRecord wording, competitors and citations
    →
    3Trace the sourceFind the page that taught the model
    →
    4Correct the originFix the source, then verify again
    ↺ Repeat monthly. Consensus compounds in both directions.

    What this actually costs in human time

    The real cost of AI search optimization is not a license fee. It is people, and hours, and the patience to let authority compound. Here is a realistic picture of what that looks like at different sizes. Treat the hours as typical, not exact, and scale them to how competitive your category is.

    Updated: July 2026

    Company size Who owns it Human hours per quarter If short on time, do only
    Startup (under 50) Founder or first marketer 20 to 30 hours The monthly audit plus one original data point
    Mid-market (50 to 500) Content or SEO lead plus one named expert 40 to 60 hours Mentions on cited pages plus author credibility
    Enterprise (500+) A GEO owner plus PR and product marketing 80+ hours Distribution and consensus of assets you already have
    Pro Tip

    Enterprises usually do not have a creation problem. They have named experts, analyst relationships, and real data sitting in silos that AI never sees. If that is you, spend your hours on distribution and consensus, not on making more.

    The Bottom Line

    For most B2B companies: pick one human job and go deep for a quarter before you buy anything. If AI gets your story wrong, that is Job 1 and Job 6. If AI ignores you, that is mentions and proof. A tool can tell you you are invisible. It cannot make you worth citing.

    For enterprises: you already have the raw material. Named experts, analyst relationships, proprietary data. The gap is almost always that it lives in places AI never reads. Your work is distribution and consensus, not another content push.

    For small teams and bootstrapped founders: you cannot outspend anyone here, which is the good news. Genuine community presence and original data are close to free, and they are exactly the things larger competitors are too busy or too cautious to do well. Speed and specificity are your edge.

    Reality check: none of this is fast, and anyone selling you instant AI visibility is selling the surface. The base takes people and months to build. That is also the reason it holds once you have built it. The companies showing up in ChatGPT today did not optimize for AI. They accumulated human trust over years, and now it is paying dividends they did not plan for. You can start accumulating yours this week.

    Frequently Asked Questions

    Do I need an AI visibility tool for this?

    For most of the checklist, no. Positioning, outreach, author credibility, community work, and original research are all human tasks that require no special software. A tool helps you measure where you stand and find which pages already cite you, which saves real time, but it does not build the trust for you. Buy one to point your effort, not to replace it.

    How is this different from regular SEO?

    SEO is largely about things you do to your own pages. AI search is shaped by what the wider web appears to agree is true about you, so the center of gravity moves off your site and onto third-party sources. The two overlap on fundamentals like authority and clear structure, but AI search puts far more weight on independent mentions, named expertise, and original data than classic ranking ever did.

    How long before AI reflects the work?

    It depends on the platform. Perplexity and Google AI Overviews pull from the live web, so changes can show up in weeks. ChatGPT leans more on training cycles, so the same change can take months to propagate. A practical approach is to start where feedback is fast, watch Perplexity and AI Overviews first, and let the slower systems catch up.

    Can I outsource the human parts?

    Some, not all. PR outreach and research production can be handled with outside help. Genuine community presence and your actual point of view cannot be convincingly faked or delegated, and trying tends to produce the astroturfing patterns that set you back. Be honest about which parts are truly yours to do and which are safe to hand off.

    What is the one thing to do this week?

    Run 20 to 30 of your real buyer questions through ChatGPT, Claude, and Perplexity, and write down exactly how you are described and who gets recommended instead. It takes an afternoon and it is the honest baseline everything else builds on. You cannot fix an answer you have never actually read.