There's a new line item appearing in marketing budgets this year, and most of the people approving it can't explain what the money actually buys. It gets called GEO services, AEO, AI search optimization, or an AI visibility audit. Underneath the labels sits one question: when a customer asks ChatGPT or Perplexity who to hire, can the machine name you?
At Scala Technologies we've spent the past year doing this work with our hands: on our own properties, on Alex Digital 360 (the WhatsApp sales automation company we operate across Latin America), and on client systems. This article is the implementation view. Not "what is GEO," which a hundred explainers already cover, but what the work actually consists of when someone does it properly: what an AI visibility audit examines, which technical tasks matter, which ones the evidence has quietly buried, and what you should expect if you pay someone to do it for you.
If you want the strategy-level story of why we started caring, our sister company published the marketing-side version with the full funnel data. This is the engineering room next door.
The mechanics: what the citation studies actually measured
Every credible GEO tactic traces back to a small pile of real research, and it's worth knowing the pile, because everything not on it is folklore.
The academic anchor is a Princeton-led study presented at KDD 2024, which tested nine optimization methods across 10,000 queries against generative engines. Its headline findings: adding statistics to a page lifted citation likelihood by 41%, quotations by 28%, and citing sources by 30 to 40%. Sites ranked fifth or lower in classic search gained up to 115% visibility inside AI answers. Translation for implementers: generative engines reward pages that are dense with liftable, attributable facts, and they reward them somewhat independently of search rank.
The industry-scale evidence comes mainly from three sources. Ahrefs correlated AI visibility across 75,000 brands and found branded web mentions (0.664) and YouTube mentions (0.737) as the strongest predictors, with backlinks far behind at 0.218. Semrush's citation studies mapped which domains each engine actually quotes and found ChatGPT citing pages that rank at Google position 21 or deeper almost 90% of the time. Profound analyzed 680 million citations and showed each engine has its own diet: Reddit is nearly half of Perplexity's top citations but only 11% of ChatGPT's.
Three engineering conclusions fall out of this data.
First, AI visibility is an entity problem, not a page problem. The models form an opinion about your business from every mention of it across the web, then select citations to justify that opinion. You optimize the opinion, not just the page.
Second, the leaderboard is not Google's. The 75% overlap between Google's top 10 and AI citations in mid 2025 collapsed to under 40% by early 2026. Whatever your SEO position, your AI position is a separate, newer, and far less contested race.
Third, the work splits cleanly into on-site tasks (quotability, structure, freshness) and off-site tasks (mentions, listicles, reviews, community presence), and the off-site half carries roughly three times the weight. Any GEO proposal that is all on-site work is charging you for the lighter half.
What an AI visibility audit actually examines
When we run one, it has five parts. If you're buying an audit from anyone, this list is your quality check.
1. Answer-space mapping. Ask each engine the questions your customers ask: best providers in the category, cost questions, comparison questions, in every market and language you operate in. Record who gets named, who gets cited, and what claims the AI makes about each. This is your actual competitive landscape, and in our experience it surprises the client roughly every time. Businesses dominant in Google turn out invisible in ChatGPT, and vice versa.
2. Entity consistency scan. Collect every significant mention of the business across the web: site, social profiles, directories, review platforms, press, partner pages. Score how consistently the name, description, service list, and location match. Language models triangulate identity from these mentions. When half the web calls you a "digital marketing agency" and half a "WhatsApp automation company," the model's internal picture is blurry, and blurry entities don't get recommended. In every audit we've run, this is where the cheapest wins live.
3. Citability review of money pages. For each revenue-relevant page: does it contain specific statistics with sources? Declarative, unhedged sentences a model can lift verbatim? Question-phrased sections matching how people actually prompt? Visible dates? The Princeton numbers put rough percentages on each of these. Most business sites fail this review not for lack of content but for lack of liftable sentences: everything reads fine and quotes terribly.
4. Third-party footprint review. Which comparison articles and listicles exist in the niche, and is the client in them? Best-of listicles account for 43.8% of ChatGPT's citations for recommendation queries, so absence from them is close to absence from the answer. Review-platform coverage matters in a stranger way: in one SaaS study, 100% of the tools ChatGPT recommended had Capterra profiles and 99% had G2 profiles, yet those platforms received under 1% of the citations. The models read reviews to form opinions, then cite blogs. Reviews are testimony, not trophies.
5. Measurement setup. AI referrals mostly arrive as direct traffic or branded search, so dashboards undercount them structurally. The fix we ship is unglamorous: a required "how did you find us" question in the intake flow. On the WhatsApp funnels we build, the qualification bot asks it as step one, and the answers are more honest than any attribution model. Plus a monthly re-run of the answer-space map, which becomes the KPI: are we named, in which answers, with what claims.
The implementation work, ranked by evidence
After the audit comes the build. Here's the work in the order the data supports, which is noticeably different from the order most GEO service packages sell it.
Entity repair first. One canonical name, one-line description, and service list, propagated everywhere: site metadata, LinkedIn, Google Business Profile, directories, review platforms, social bios, partner pages. Tedious, cheap, high-leverage. This is plumbing work, and like most plumbing it's invisible until you see the before-and-after in the answer maps.
Comparison content second. Getting the client into existing listicles where possible, and building honest comparison pages where they don't exist. The counterintuitive part: engines cite comparison pages even when a participant publishes them. The most-cited page we've ever shipped for our sister brand is a detailed chatbot comparison, exactly the format the citation data predicts, and it outperforms every services page on the domain for AI visibility.
Quotability rewrites third. Statistics with sources, definitive sentences, FAQ blocks answering literal question phrasings, visible dates, quarterly refresh cycles. Content cited by ChatGPT averages 393 days newer than what ranks in Google for the same query, so the refresh cycle isn't optional polish. It's a ranking factor by another name.
Off-site presence fourth, and continuously. Earned mentions in industry roundups, podcast appearances, useful answers in the Reddit threads where the client's customers ask for recommendations, YouTube presence where it fits. This is the three-to-one half. It's also the half that can't be faked at scale: astroturfed community presence gets detected by both the platforms and, increasingly, the models. Slow and genuine is the only version that compounds.
Structured data last, as hygiene. We keep schema clean because Google says structured data feeds its own AI surfaces, and because it costs little. But no citation study has found schema coverage correlating with ChatGPT or Perplexity citations, so we position it honestly: hygiene, not strategy. Anyone selling "schema for ChatGPT" as a headline deliverable is selling the audit-proof part of the work.
The technical baseline: crawler access and retrieval paths
Before any of the content work matters, the machines have to be able to read you, and there's a short technical checklist that a surprising number of businesses fail without knowing it.
Modern AI answers reach your site through two distinct paths, and they have different user agents. Training crawlers (GPTBot for OpenAI, Google-Extended for Gemini training, ClaudeBot for Anthropic) feed the models' background knowledge of your entity. Retrieval crawlers (OAI-SearchBot for ChatGPT's live search, PerplexityBot, Google's ordinary Googlebot for AI Overviews) fetch pages at answer time, and they're the ones behind live citations. A robots.txt written years ago by a cautious developer, or a firewall rule that rate-limits unfamiliar bots, can block either path silently. We've audited businesses investing in content while their WAF served the retrieval crawlers a 403 on every request. Check your server logs for these user agents before spending a peso on anything else; their absence is a finding in itself.
Two related items round out the baseline. Rendering: retrieval bots deal poorly with content that only exists after heavy client-side JavaScript, so revenue-relevant claims belong in server-rendered HTML. And response time matters at answer time in a way classic SEO never demanded: a retrieval bot assembling an answer in seconds doesn't wait patiently for a slow origin. None of this is exotic engineering. It's twenty minutes of log reading and a robots.txt review, and it gates everything downstream.
What an engagement looks like in practice: a walkthrough
An anonymized composite from our audit work, because the abstract list above lands differently with numbers attached.
A LATAM services business, healthy Google rankings, came in asking why a competitor kept being named by ChatGPT while they were invisible. The answer-space map confirmed it: across twenty customer-shaped questions in two languages, the competitor was named in fourteen answers, the client in one. The entity scan found the usual suspect: four different self-descriptions across their site, LinkedIn, directories, and an old press page, including a legacy brand name from before a rebrand. The citability review found beautifully written pages with almost no liftable sentences and no visible dates. The footprint review found the competitor present in three industry listicles the client had never heard of, and the client's review profiles dormant since 2024.
The first month was entity repair and quotability rewrites on five money pages. The second was outreach that landed them in two of those three listicles plus one new comparison piece. By the third monthly answer-map re-run, they were named in six of the twenty answers; by month five, eleven, including the two highest-intent cost questions in Spanish, where they were frequently the only business named at all. Total engineering effort was modest. The work was never clever, it was complete, and that's the honest shape of this discipline: the wins come from doing unglamorous things all the way through, in a field where most competitors haven't started.
What we tested and dropped
Engineering credibility comes from the negative results, so here are ours.
llms.txt. We implemented it early across our properties. Nothing moved, and the ecosystem data explains why: Google's John Mueller confirmed no AI service uses it, and Ahrefs found 97% of llms.txt files across 137,000 domains received zero requests, ever. It remains on our servers as a monument to doing what the checklist said. If a proposal features llms.txt prominently, read the rest of it with suspicion.
Content volume. Publishing more pages correlates with AI visibility at 0.194, which is noise. We rebuilt our editorial cadence around fewer, denser, citable pieces and saw more movement than any volume push produced.
Multi-engine optimization from day one. Each engine's citation diet is different enough that chasing all of them dilutes the work. ChatGPT drives about 87% of AI referral traffic, so for most clients we optimize for ChatGPT first, treat AI Overviews as the SEO-adjacent second, and touch Perplexity only where the client's audience actually lives on Reddit.
Why this matters commercially: the conversion math
The traffic volumes are still small, and any honest implementer says so out loud. What makes the work pay is intent density. Semrush found AI search visitors converting at 4.4 times the rate of organic visitors; Seer Interactive's data puts ChatGPT referrals around 16% conversion against under 2% for Google organic. The counterevidence is real too: for ecommerce, several studies show AI referrals converting no better or worse than organic. The pattern is consistent though: for services and B2B, where the AI acts as a pre-sales consultant, the referral arrives pre-convinced.
We see this end to end because we build the receiving side too. In the WhatsApp funnels we engineer for LATAM businesses, AI-referred leads complete qualification flows at roughly double the rate of paid-traffic leads. The AI already did the comparison shopping, the objection handling, and the shortlisting inside a conversation nobody can see. What arrives is a customer doing paperwork. For the systems-level view of what happens to those leads next, the $500 a month sales stack breakdown on our sister site shows the full pipeline.
And the horizon item we're building toward: agentic commerce, where a customer's AI doesn't just recommend but contacts the shortlist directly. When that message lands, it lands on a channel. In Latin America that channel is WhatsApp, which is why we treat AI visibility engineering and conversational automation as one system with two ends: get named by the machine, then receive the conversation the naming produces, instantly, in the customer's language, at any hour.
What GEO services cost in 2026
Pricing in this field is chaotic because the field is new, so here's an honest map of what we see in the market.
Standalone AI visibility audits run from a few hundred dollars for automated tool reports to $2,000 to $5,000 for the manual, answer-map-driven version described above. Monthly GEO retainers at agencies typically land between $1,500 and $5,000 for small and mid-size businesses, scaling with how contested the niche is and how much off-site outreach the plan includes. Enterprise engagements with multi-market, multi-language scope go well beyond that. The tooling layer, AI citation trackers and brand-monitoring platforms, adds $100 to $500 a month if you want continuous measurement rather than monthly manual maps.
Calibrate against the work: the on-site half is front-loaded and shouldn't recur at full price forever, while the off-site mention flywheel is genuinely ongoing. A retainer that's all "content optimization" after month three is billing you for the light half. And since there's no paid placement to buy, every dollar goes to earned position, which is precisely why early spend in an uncontested niche compounds: the position you engineer now is the position a competitor must dislodge later.
Does this replace your SEO investment?
No, and any pitch framed as "SEO is dead" should end the meeting. Google still carries over 90% of searches, and AI Overviews are partially assembled from Google's own index, so classic rankings still feed one of the three big answer surfaces. What has changed is the payout curve: Pew's behavioral data shows users clicking results roughly half as often when an AI answer is present, and Ahrefs measured top-position clickthrough losses up to 58% under AI Overviews. Your rankings are keeping their positions and losing their clicks.
The practical budget answer is reallocation, not replacement. Keep the SEO fundamentals: crawlability, site speed, the content that ranks. Redirect the marginal dollar, the one that used to chase one more backlink or one more keyword post, toward the GEO stack: entity work, comparison content, earned mentions, quotability. Almost everything in the GEO stack strengthens SEO as a side effect, while the reverse is only partially true. That asymmetry is the whole argument.
The LATAM asymmetry
One market observation we'd be negligent not to flag. The Spanish-language answer space is dramatically thinner than the English one. Ask ChatGPT in Spanish for the best providers in almost any Colombian, Mexican, or Chilean niche and the answers draw from generic, Spain-centric content, because almost no citable Spanish comparison content exists for LATAM markets.
For implementers, this is the cheapest arbitrage in the field right now. The same playbook that fights for scraps in English competes against nearly nobody in Spanish. A LATAM business that builds citable Spanish comparison content, consistent bilingual entity data, and a handful of earned mentions in the next year becomes the default recommendation in its niche more or less by walkover. We've watched it happen in our own market. The window is measured in months, not years.
What to expect if you hire someone for this
Whether it's us or anyone else, a serious GEO engagement looks like this, and deviations are diagnostic.
- A baseline answer-space map before any work starts. If there's no before, the after is unfalsifiable.
- Entity and citability fixes in the first month. They're cheap and they move first.
- An off-site mention plan with named targets. Which listicles, which roundups, which communities. "We'll build brand awareness" is not a plan.
- Monthly re-runs of the answer map as the reporting artifact. Are you named, in which answers, with what claims, in which languages. Screenshot-level evidence, not dashboard abstractions.
- Honest volume expectations. AI referral traffic will be a trickle next to Google. The pitch is intent density and early-mover position on a new leaderboard, not a traffic flood.
- No llms.txt theater, no schema miracles, no page-volume quotas. The evidence buried all three.
Timeline-wise: entity and on-site work shows in answer maps within one to three months in low-competition niches, slower in contested English-language categories. Off-site compounding is a six-to-twelve-month curve. Anyone promising "#1 in ChatGPT in 30 days" is selling something the mechanics don't support: there is no paid placement in these answers to shortcut with, which is precisely why the earned position is worth engineering.
Frequently asked questions
What is an AI visibility audit?
A structured review of how AI engines currently see a business: which answers name it, how consistent its entity data is across the web, how citable its key pages are, what third-party footprint (listicles, reviews, community mentions) exists, and how AI-driven customers will be measured. It ends with a baseline answer-space map that later work is judged against.
What do GEO services actually include?
Legitimate engagements cover entity consistency repair, comparison and citation-worthy content, quotability rewrites of key pages with sourced statistics and FAQ structure, an earned-mention plan targeting listicles and communities, structured data hygiene, and monthly answer-space reporting. Be wary of packages centered on llms.txt files, schema as a growth lever, or bulk content production; citation studies support none of those.
How is GEO different from SEO technically?
SEO optimizes pages for a ranked list using links and on-page authority. GEO optimizes an entity for inclusion in a synthesized answer, where brand mentions across the web outweigh backlinks roughly three to one, freshness runs about 400 days ahead of Google's preferences, and your Google position barely predicts your citation odds. The two share groundwork but are scored by different machines reading different signals.
Can you pay to appear in ChatGPT results?
No. There is no ad placement inside ChatGPT's organic answers, which is exactly why earned visibility there converts so well: the recommendation reads as neutral because it is. The only route is being the entity the model finds consistently described, frequently mentioned, and citably documented.
How long does it take to show up in AI search results?
In thin niches and non-English markets, one to three months for entity and on-site fixes to surface in answer maps. Competitive English categories run longer, and the off-site mention flywheel compounds over six to twelve months. Freshness signals mean results also decay without maintenance, so it's a rhythm, not a project.
How do you measure AI search visibility?
Two instruments. Monthly answer-space maps: the same customer-shaped questions run against each engine, recording who is named and with what claims. And first-party attribution: a required "how did you find us" question in the intake or qualification flow, because AI referrals hide inside direct and branded-search traffic in every analytics dashboard.
Does schema markup help with ChatGPT visibility?
Not measurably. Citation studies have found no correlation between schema coverage and LLM citations, and the Princeton study never tested it. Keep structured data clean for Google's AI surfaces, where Google's own guidance says it helps, and treat it as hygiene rather than a growth lever.
How much do GEO services cost?
Manual AI visibility audits run $2,000 to $5,000; monthly GEO retainers for small and mid-size businesses typically land between $1,500 and $5,000 depending on niche competitiveness and how much off-site outreach is included; citation-tracking tools add $100 to $500 a month. On-site work is front-loaded, so a healthy engagement gets cheaper or shifts toward earned mentions after the first quarter rather than billing flat forever.
Should I block AI crawlers from my site?
For a business that wants customers, almost never. Blocking GPTBot or PerplexityBot in robots.txt removes you from the models' knowledge and from live retrieval, which in answer-driven discovery means removal from the shelf. The publishers blocking AI crawlers are monetizing content itself; a services business monetizes being found. Audit your robots.txt and firewall for accidental blocks; we find them regularly.
Can my existing marketing team do GEO in-house?
The on-site half, usually yes: entity cleanup, quotability rewrites, FAQ structure, and refresh cycles are ordinary content operations once someone has read the citation research. The off-site half (earning listicle placements, community presence) and the technical baseline (crawler access, rendering, answer-map tooling) are where outside help tends to pay. A sensible split is an external audit and baseline, then in-house execution against the map with quarterly external re-runs.
Is GEO worth it for small businesses?
Often more than for large ones. AI answers in local and niche categories are assembled from thin source material, so a single well-engineered comparison page, consistent entity data, and a few earned mentions can make a small business the default recommendation, especially in Spanish-language LATAM markets where citable content barely exists.
The bottom line
Getting recommended by AI is an engineering problem with a marketing surface. The inputs are unglamorous: consistent entity data, citable sentences, comparison content, earned mentions, fresh dates, honest measurement. The evidence for each is public, and so is the evidence against the theater versions. If you want the work done on your business, from the audit through the WhatsApp funnel that catches what the AI sends, that's what we build at Scala Technologies. And if you'd rather run it yourself, everything above is the checklist. Either way, run the answer-space map this week. Finding out what the machines currently say about you costs nothing, and it's rarely what you expect.
