Ask ChatGPT to recommend a vendor in your category. Then ask where it got its answer. The sources are rarely the company websites. They’re a roundup from two years ago, a Reddit thread, a review site and maybe one trade article. Your own homepage is usually the least important page in the answer.
That’s the problem every GEO buyer runs into, and it matters more each quarter. Forrester’s 2026 predictions found that 30% of buyers in 2025 considered generative AI tools a meaningful interaction at the final purchase stage. Only 17% said the same about talking to product experts. So a chatbot is now sitting in on the purchase decision, and it’s quoting other people about you.
Generative engine optimization services promise to change what it says. Some do. Plenty just resell SEO with a new cover page. This guide explains how to tell the two apart before you sign.
Why AI answers are getting the budget
The audience is huge. OpenAI announced in February 2026 that ChatGPT had passed 900 million weekly active users. Google has turned into an answer engine too. In July 2026, Search Engine Roundtable reported that Similarweb measured AI Overviews on about 43% of Google queries, and Semrush on about 48%.
Clicks are also falling. Ahrefs compared 300,000 keywords from December 2023 to December 2025. It found AI Overviews linked to a 58% lower click-through rate for the page ranked first. So even when you rank first, the AI summary above you answers the question and keeps the click.
That leaves a simple question for buyers: when an AI answers for you, what is it reading?
The source list behind every AI answer
This example is made up, but the pattern is common. A reputation firm for law practices runs the prompt “best reputation management company for a law firm” through Gemini 20 times. It shows up in three of those answers. A competitor shows up in 17. The citations explain why. Gemini keeps pulling from a 2023 “top 10 vendors” list that left the firm out, plus a forum thread where an unhappy former client vented. ChatGPT, meanwhile, lists the firm’s old office address, which it picked up from an outdated directory profile.
A blog post can’t fix any of that. The model already has its sources, and the firm’s website isn’t one it leans on.
Real GEO work starts from that fact. AI models give a lot of weight to what independent sources say about a brand, and much less to what the brand says about itself. If an agency can’t show you which pages feed answers in your category, it’s working blind.
What GEO services actually include
Most legitimate engagements are built from six kinds of work. Few firms do all six well, so ask about each one.
| Workstream | What you should receive | Does it change what AI quotes? |
| Prompt research and tracking | A prompt list, plus baseline mention rate and share of voice for each model | No. It measures. |
| Citation and source mapping | The domains and pages behind answers in your category | It tells you where to work |
| Off-site mention building | Earned media, reviews, roundup inclusion | Yes, directly |
| Entity cleanup | Consistent company facts across profiles, plus schema markup | Yes, for accuracy |
| Content restructuring | Your top pages rewritten so passages are easy to quote | Somewhat |
| Crawler access | Fixes to robots.txt and JavaScript rendering | Yes. Without it, nothing gets read. |
Prompts and tracking come first
A prompt is how a real buyer asks: “Which payroll tool works for a 40-person restaurant group?” A keyword like “payroll software” won’t tell you much about AI answers. Good firms run each prompt many times on several models, because answers change from run to run. A single screenshot of one AI answer proves nothing.
Source mapping is the deliverable weak vendors skip
The source map turns tracking into a plan. If Perplexity keeps citing one comparison page, the job is getting onto that page or onto the one that replaces it. If a Reddit complaint keeps coming back, that thread is the job.
Off-site work should eat most of the hours
Here’s where real GEO services spend most of their hours: on sites the client doesn’t own. That means expert commentary in the publications your source map flagged, honest review requests to real customers, and outreach to the editors of “best X for Y” lists. It also includes disclosed, useful posts in the forums where buyers compare options, and accurate profiles in industry directories.
If a proposal is 80% blog posts, you’re buying content marketing. That has value. It just doesn’t touch the sources the model already trusts.
Entity cleanup stops the model from guessing
When your founding year, leadership, or service list differ across LinkedIn, Crunchbase, Google Business Profile, and old directories, models fill the gaps. Sometimes they merge you with a company that has a similar name. Cleanup makes those facts match everywhere. It also adds Organization and Person schema with sameAs links to your official profiles. Schema helps machines identify you correctly. It won’t get you recommended on its own.
Crawler access is the cheapest win
Some sites block GPTBot, ClaudeBot, or PerplexityBot in robots.txt without knowing it. Others load key text through client-side JavaScript, which many AI crawlers don’t run. A competent firm checks both in week one.
What Google says you can skip
In May 2026, Google published official guidance for its AI features. It’s blunt. You don’t need llms.txt or other “AI files.” You don’t need to chop content into small chunks. And there’s no special schema markup for AI Overviews or AI Mode. Google’s AI features run on the same index and ranking systems as regular search, which is why the basics in our breakdown of Google ranking factors in 2026 still decide what gets cited.
That doesn’t make clear writing worthless. A direct answer near the top of a page, headers phrased like real questions, and named sources for your numbers all help readers. They also help any system that quotes short passages. Just treat them as good editing. An agency billing you separately for “llms.txt implementation” or “AI chunking” is selling something Google says it doesn’t need.
A realistic first 90 days
Month one: the prompt set, baseline data across at least three models, the source map, a crawler check, and a list of false claims AI tools currently make about you.
Month two: entity cleanup mostly done, priority pages rewritten, and outreach underway against a named list of targets.
Month three: the first earned placements, a second round of visibility data compared against the baseline, and a clear view of which models are moving.
Don’t expect a big jump in share of voice by day 90. New third-party coverage has to be crawled and worked into answers first, so bigger shifts usually show up somewhere between months four and nine.
What GEO costs in 2026
WebFX’s October 2026 pricing guide puts agency retainers at:
- Small businesses: $1,500 to $5,000 a month
- Mid-sized companies: $5,000 to $25,000 a month
- Enterprises: $25,000 to $50,000+ a month
Standalone AI visibility tracking software runs from about $50 to over $1,000 a month, depending on how many prompts and platforms you track. One-time audits vary a lot, from a few thousand dollars into five figures.
How the money is split matters more than the total. Ask what share of the hours goes to off-site work. A $4,000 retainer with real outreach can beat a $15,000 one that’s mostly dashboards and blog posts.
Reading a GEO proposal
Check these before you compare prices:
- Does it list real prompts for your category?
- Which models does it track, and how many times is each prompt run?
- What percentage of hours goes to sites you don’t own?
- Does it name the publications, review sites or lists it will go after?
- Will you get a measured baseline before work starts?
- Does it check whether AI describes you correctly, not just whether it names you?
Walk away from anyone who guarantees placement in ChatGPT or Google’s AI answers. Nobody controls what these models say.
Where GEO hits a wall
GEO can’t make a weak reputation look strong. If the sources models trust are full of complaints, lawsuits, or bad press, the models will repeat them. An agency can add better material to the mix, but it can’t delete what’s true and widely covered. That’s why GEO and reputation management increasingly overlap. Both start with the same question: what does the web already say about you, and which of those sources carry weight?
It also can’t replace traditional search. Pages that rank well in Google still feed a large share of AI citations. If your SEO has slipped, GEO has a weak foundation. Our piece on brand visibility in AI search covers the free reporting tools from Google and Bing that show you where you stand before you hire anyone.
Is AI traffic worth paying for?
There isn’t much of it, and the evidence on quality is mixed. In June 2025, Ahrefs reported that AI search drove 0.5% of its visits over 30 days but 12.1% of signups. That’s about 23 times the conversion rate of regular search visitors. Semrush estimates an AI search visitor is worth 4.4 times a standard organic visitor. Then there’s the other side: a study of 973 ecommerce sites by Kaiser and Schulze found ChatGPT referrals converting worse than Google organic search.
A fair reading is that B2B and high-consideration purchases look strong, while average ecommerce looks weaker. Measure your own. Tag AI referrers in analytics, add “AI assistant” to your “how did you hear about us” field, and watch branded search volume. A lot of GEO’s value never shows up as a click. It shows up as a prospect who already knows your name when they call.
Doing it in-house or hiring out
A capable marketing team can handle prompt tracking, robots.txt fixes, schema and page rewrites. The hard part is off-site. Earning placements in the publications models trust takes media relationships, editorial judgment and patience most internal teams don’t have.
Many mid-sized companies split the work. They keep tracking and on-site fixes in-house and hire out earned media and review growth. Either way, keep the prompt set and baseline data under your own control, so you can switch vendors without starting over.
Quick answers
Q.1 Can a GEO agency remove a bad AI answer?
No. AI answers can’t be deleted on request. An agency can correct or crowd out the sources behind them, and the answers change as models re-read those sources.
Q.2 Does llms.txt help?
Not for Google. Its May 2026 guidance says AI features don’t need special files or markup. Crawler access and real third-party coverage matter far more.
Q.3 Is GEO separate from SEO?
Google calls it SEO. In practice, the weighting changes: third-party mentions, consistent company facts, and repeated prompt testing count for more than most SEO plans assume.
One last test before you sign: ask the agency to show you which sources AI tools cite in your category today. If they can’t answer before you hire them, they probably won’t be able to after.
Related: Google Cloud vs. On-Premises for AI: Which Costs Less in 2026?
