ChatGPT Ads and GEO: How Paid and Earned AI Visibility Fit

24 September 2026 · The Cresia team

Paid and earned AI visibility are different channels that happen to share an interface. ChatGPT ads let a brand pay to be present in the conversation, while GEO is the slower work of being named in the answer because the model has good reason to name you. Fund and measure them separately, and do the earned work first, because ads cannot fix an answer that describes you badly.

Key takeaways

ChatGPT ads and GEO are separate channels: one is bought, the other is earned by being a source the model trusts.
Paying for placement does not change what the model says about you, so the earned work still has to be done.
Answer engines favour pages that state a claim plainly, agree with other sources and are easy to fetch.
Give paid and earned AI visibility different owners and different success metrics before spending anything.
Measure with a fixed prompt set, logged over time, and treat any single screenshot as anecdote.
Ignore promises of guaranteed citations and one-off tricks; fix your content, entity data and tracking.

What ChatGPT ads change for AEO and GEO teams

Search Engine Journal reported a discussion between OpenAI and Go Fish Digital about ChatGPT ads, earned AI visibility and how to measure GEO in practice. The useful part of that pairing is the framing itself: an AI assistant now has a paid layer and an earned layer, and a marketing team has to decide what belongs in each.

That sounds obvious until you look at how most organisations are set up. The SEO lead owns citations in AI answers, if anyone does. The paid media lead owns anything with an invoice. Nobody owns the question of what happens when both show up in the same conversation, and the two teams read different dashboards with different definitions of a win.

The shift is small in mechanics and large in organisation. Before, a brand's presence in ChatGPT was entirely earned, and it was hard to buy your way in. Now there is a bought route, and the temptation will be to treat it as a shortcut around the messy work of being cited.

It isn't a shortcut. An ad and an answer are different objects. The answer is the model's synthesis of what it can find about a topic. The ad is your message, labelled as yours. A buyer who reads both will notice if they disagree.

Formats, targeting, eligibility and pricing for ChatGPT ads are OpenAI's to define and they are likely to change. Read OpenAI's own advertising documentation and your account dashboard for the current details rather than building a plan on a summary of a summary.

Why paid and earned AI visibility are not the same job

Think of a mid-size B2B software company selling contract management. A buyer asks an assistant which tools suit a legal team of ten. The earned question is whether the assistant names the company in its answer, and how it describes it. The paid question is whether the company's own message appears alongside.

The two have different failure modes. Earned visibility fails when the model doesn't know you, confuses you with another product, or repeats a stale description from an old review site. Paid visibility fails when the message is generic, the landing page doesn't match what the buyer just read, or the spend goes to prompts nobody in your market actually asks.

They also move at different speeds. You can start or stop a paid campaign in an afternoon. Earned visibility is the residue of years of published pages, third-party mentions and consistent facts. You can improve it in weeks, but you can't switch it on.

That asymmetry is the argument for sequencing. If the answer already describes you accurately and favourably, ads amplify a good position. If it doesn't, ads pay to sit next to a description you'd rather change. Fix the description, then spend.

One more difference matters for reporting. Paid media reports arrive with impressions and costs. Earned AI visibility arrives as observations: on this date, for this prompt, the assistant said this. Mixing the two in a single number produces something that looks precise and means little.

How answer engines choose and cite sources

No assistant publishes its full selection logic, and the behaviour changes between products and between model updates. What you can say safely is limited, but it's enough to act on.

Answer engines generally combine what the model learned in training with material fetched at question time. The fetched part comes from a search index or a crawler, and the model then writes an answer and, in many products, attaches links. Whether a page gets fetched depends on whether it can be reached, rendered and matched to the question.

From that, four practical things follow.

Reachability. A page blocked from the relevant crawler, hidden behind a script that never renders, or slow enough to time out can't be cited. This is the cheapest failure to fix and the most commonly missed.
Match to the question. Models retrieve passages, not whole sites. A section that answers one question completely is easier to lift than a long page that circles it.
Agreement across sources. A claim that appears consistently on your site, in your documentation and on independent pages is safer for a model to repeat than one that appears only on your homepage.
Clarity about who you are. Names, product descriptions, pricing models and categories that vary from page to page invite the model to blend you with something else.

Search Engine Roundtable covered a report that Google AI Overviews now carry many more external links than before. If that holds, it means more places where a source can be named, and more competition to be one of them. It also means citation counts alone tell you little without knowing which pages were cited and for which questions.

For Google's own features, its Search Central blog and documentation are the place to confirm what it says about content eligibility. For ChatGPT, use OpenAI's help pages. Don't rely on a consultant's claim about a ranking factor unless they can point to a vendor document.

Who owns what: a paid and earned split

The most useful thing a team can do this month is write down who owns each part. A table on one page beats a strategy deck nobody rereads.

Area	Paid (ChatGPT ads)	Earned (GEO)	Suggested owner
Goal	Reach defined audiences in the conversation	Be named and described accurately in answers	Paid lead / SEO lead
Inputs	Creative, targeting, landing pages, budget	Content, entity facts, third-party mentions, crawl access	Media team / content team
Speed	Days to launch or stop	Weeks to months to shift	Both
Main risk	Wasted spend, message mismatch	Stale or wrong descriptions persisting	Paid lead / SEO lead
Evidence	Platform reporting, click and conversion data	Logged answers to a fixed prompt set	Analytics lead
Review cadence	Weekly	Monthly	Growth lead

Two rows deserve comment. The analytics owner is separate because neither the paid nor the SEO lead should grade their own work. And the review cadence differs on purpose: checking earned visibility weekly mostly measures noise from the model's own variation.

If your organisation runs media and search under one roof, the media teams view of the platform is a reasonable starting point for how a shared workflow can look. If they're separate, put a fifteen-minute monthly call on the calendar where both leads read the same prompt log.

What to change in practice

Here is a sequence that works for a team of any size. It assumes you have a site, an analytics setup and at least a few pages that already earn organic traffic.

Check crawl access. Read your robots rules and any bot protection settings. Make sure the crawlers you want, including those documented by OpenAI and Google, can fetch your key pages. Do it before anything else.
Write a fact sheet for your own brand. One internal page listing your official name, category, products, pricing model, target customer and three things you are not. Every public page should agree with it.
Rewrite your ten most important pages so a passage stands alone. Put the answer in the first two sentences under each heading. Name the product and the audience in the sentence, not just in the heading above it.
Fix old descriptions off-site. Review listings, partner pages, directory entries and old press mentions. A wrong description on a page you don't control is often what the assistant repeats.
Build comparison and alternatives content you can defend. Buyers ask assistants which option suits their situation. A page that states honestly who you fit and who you don't is quotable. A page that claims to fit everyone isn't.
Then decide on paid. Choose a small set of prompts where you already appear accurately, write ad copy that matches the answer's language, and send clicks to a page that continues the same thread.

Step six is where many teams will rush. A landing page built for a search keyword often fails for an assistant visitor, who arrives having already read a paragraph about the category. They don't need the category explained again. They need to know whether you fit their situation.

Creative variation is another place to be careful. If you plan to test several messages, a structured A/B testing approach keeps you from concluding anything from two days of traffic. That applies to landing pages more than to the ads themselves.

How to measure paid and earned AI visibility

Measurement is where this topic goes wrong most often, so be strict about it.

Start with a fixed prompt set. Write 30 to 60 questions your buyers actually ask, grouped by stage: category discovery, comparison, and pre-purchase checks. Keep the wording stable. Run them on a schedule, in a consistent way, and log the full answer, the date, the product and any brands and links named.

Assistants vary between runs. The same prompt can produce different answers an hour apart. So a single observation proves nothing. What you're looking for is direction across the set over several weeks: does your name appear more often, is the description closer to your fact sheet, are the linked pages the ones you'd choose?

For referral traffic, tag your paid landing links consistently and check how your analytics platform classifies visits from assistants. Classification is often imperfect, and some visits arrive with no referrer at all. Expect undercounting of earned visits and don't be surprised when direct traffic quietly rises.

A tracking specification helps here. If campaign parameters, event names and landing page groupings are written down and enforced, paid and earned visits can be compared without argument about definitions. Without one, you'll spend the first meeting reconciling spreadsheets.

Tools that log AI answers at scale exist, and platforms including MediaPilot address AI-search visibility, so look at what the product page says and test it against your own prompt set. Whatever tool you choose, ask how it runs prompts, how often, and how it handles variation between runs.

A simple starting metric set:

Share of prompts where your brand is named, tracked as a trend, not a target.
Accuracy of the description against your fact sheet, scored by a person.
Which of your pages are linked, and whether they're the ones you'd choose.
For paid: cost per qualified visit and conversion rate on the matched landing page.

Notice that none of these is a rank. There is no stable position one in an answer, and any report that presents one should be read sceptically.

What not to do

The first mistake is buying visibility to cover a content problem. If an assistant describes your product as something you retired two years ago, an ad doesn't correct the record. The next buyer still sees the old description.

The second is chasing a trick. Every few weeks someone claims a special file, a phrase or a schema type that forces citations. Treat these with suspicion unless a vendor's own documentation supports them. Structured data and clear markup help machines read a page, but they don't override a lack of substance.

The third is reporting a single number to the board. A dashboard tile reading 34 per cent share of voice hides the prompt set, the run date and the variation. Give leadership the trend, the method and the caveats.

A few more habits to drop:

Stuffing pages with question headings that have thin answers underneath.
Publishing dozens of near-identical comparison pages for every competitor pair.
Judging success by whether a hand-picked prompt shows your name today.
Setting one team's bonus on a metric that the other team's work moves.

There's also a sensible stance on volume. Publishing more pages rarely helps if the existing ones contradict each other. A tidy set of forty accurate pages beats four hundred loosely edited ones, both for readers and for any system that has to decide which of your pages to trust.

What is still uncertain

Several things could shift within months. Ad formats and placement in ChatGPT are new and likely to change. How closely paid and organic answers are separated in the interface is OpenAI's decision. Google's own AI features, and the merchant and checkout changes reported by Search Engine Roundtable, point the same way: assistants are moving from describing options toward helping buyers act on them.

That raises the stakes on accurate product data and clean landing paths, and it makes the boundary between advertising, commerce and content less tidy. Plan for change by keeping the parts you control in good order: your facts, your pages, your tracking.

On the evidence so far, a reasonable position is this. Earned AI visibility is the foundation, paid is an accelerator worth small, controlled tests, and neither has settled measurement. Teams that write down their definitions now will be able to tell what worked when the platforms mature. Teams that don't will be arguing about screenshots.

If your growth, media and analytics leads are working from different definitions, the growth teams page shows how Cresia frames shared work across those roles. The real change, though, is a meeting and a document, and neither needs software.

Frequently asked questions

Do ChatGPT ads replace GEO?

No. Ads buy presence in the conversation, but the answer the model writes still depends on what it can find and trust about you. If that answer is wrong or missing, paid placement sits next to it and doesn't repair it.

Should we start with paid or earned AI visibility?

Start with earned, because the work is cheaper to begin and it improves the context any ad appears in. Then run small paid tests on prompts where your brand is already described accurately, so you learn what the channel does without paying to amplify a problem.

How do we measure GEO if answers change every time?

Use a fixed set of prompts, run them on a regular schedule and log every answer. Judge direction across weeks, not any single result. Score accuracy against your own fact sheet, and treat share of mentions as a trend, not a rank.

Who should own AI visibility inside the company?

Split it. The paid lead owns ChatGPT ad spend and results, the SEO or content lead owns earned visibility, and an analytics owner keeps the prompt log and definitions so neither team grades its own work. A monthly review with all three keeps the picture consistent.

Can structured data guarantee citations in AI answers?

No vendor documentation we can point to says it does. Structured data helps machines read a page and it's worth doing properly, but citation depends on reachability, relevance and agreement across sources. Check each vendor's own guidance for current statements.

Sources

https://www.searchenginejournal.com/chatgpt-ads-and-geo-where-paid-and-earned-ai-visibility-fit-together/590537/
https://www.seroundtable.com/google-ai-overviews-more-external-links-42135.html
https://developers.google.com/search/blog