We Compared ChatGPT Ads to Google and Meta, Setting by Setting

Spend an hour inside OpenAI Ads Manager and you’ll have the same reaction most paid media buyers have: this looks exactly like Google Ads, and it behaves nothing like it.

The hierarchy is Google’s. The auction is Google’s. The creative is text-first, like Google’s. Then you go looking for the keyword field and there isn’t one. You go looking for negatives and there aren’t any. You open the audience library and find it empty unless you’ve uploaded something yourself.

The short version is this: ChatGPT Ads has Google’s skeleton and Meta’s nervous system, and it’s thinner than both. Here’s what that means at the level of individual settings — and where it changes how you’d actually run an account.

The quick verdict

	Google Ads	Meta Ads	ChatGPT Ads
Structure	Campaign → ad group → ad	Campaign → ad set → ad	Campaign → ad group → ad
How you target	Keywords with match types	Audiences and interests, or broad	Plain-language context hints
Query/placement visibility	Search terms report	Placement and breakdown reports	None
Negative controls	Negative keywords, placement exclusions	Placement and audience exclusions	Audience exclusions only
Audience sources	Customer Match, remarketing, in-market, affinity	Uploads, pixel, lookalikes, interests	Uploaded lists only
Main optimization lever	Query and bid refinement	Creative iteration	Creative iteration
Scheduling	Dayparting	Dayparting	Dates only
Benchmarks available	Extensive	Extensive	None published

Structure: nearly identical to Google

Campaign → ad group → ad, with the ad group as the thematic unit holding the bid and the targeting instruction. If you’ve built a search account, you can build a ChatGPT campaign in twenty minutes without reading a guide.

Campaign level holds the title, objective, conversion event, budget, dates, location, platforms and custom audiences. Ad group level holds the title, context hints, maximum bid and audience bid adjustments. Ad level holds title, copy, image, landing page URL and your advertiser name and logo.

The one structural quirk worth flagging: most targeting sits at campaign level, not ad group level. Location, platform and audience inclusion or exclusion are all campaign settings. Your ad groups only differ by theme, bid and hints. That pushes you toward more campaigns and fewer ad groups per campaign than you’d normally build on Google — one campaign per market, per product line, per audience posture.

Targeting: much closer to Meta than Google

This is the fault line, and it’s where most Google-trained advertisers waste their first month.

On Google, a keyword is an instruction. You say [emergency plumber london] in exact match and you have made a binding statement about when you appear. On Meta, detailed targeting is closer to a suggestion the system takes under advisement, and broad targeting is barely even that.

ChatGPT Ads sits on the Meta end. Instead of keywords, ad groups carry context hints — plain-language descriptions of the situations, needs and conversation types where your offer belongs. OpenAI is explicit that hints are not exact-match controls, not audience rules, and not a guarantee your ad appears for any particular topic or phrase. The system reads the live conversation, reads your hints, reads your ad and landing page, and decides whether you’re relevant.

Writing them well is a different craft to writing keywords. “Running shoes” is a weak hint. “Cushioned everyday running shoes for beginners training for their first 5K” is a strong one, because it tells the system who the product helps and when it’s useful — information not already sitting in your ad copy.

The discipline that transfers from Google: one ad group, one intent. If two hints would need different ad copy or different landing pages, they’re two ad groups.

Location is the one place ChatGPT is genuinely comparable to the big two. Country targeting is standard, and in the US you also get state, DMA and ZIP code targeting. Outside the US it varies by market — search the campaign location picker or download OpenAI’s location catalogue as a CSV.

Platforms are a multi-select for iOS app, Android app and Web (desktop and mobile web together). Include or exclude only. Compare that to Google’s device bid adjustments and Meta’s placement-level control and it’s clearly a first pass.

Bidding and budgets: Google’s model, fewer options

Three objectives, and they set both pricing and optimisation:

• CPM — pay per thousand impressions, optimised for reach.
• CPC — pay per valid click, optimised toward users likely to click.
• oCPC — conversion-optimised clicks. You pay per click, not per conversion, but delivery optimises toward a tracked event.

That’s roughly Google’s manual CPC and Maximise Conversions territory, minus target CPA, target ROAS, target impression share and portfolio strategies. Two things to warn clients about: the objective can’t be changed after creation, and some budget-type switches are one-way — move a campaign-total budget to a daily budget and you can’t move it back.

Budgets work as you’d expect. A campaign-total budget is a hard ceiling for the flight; a daily budget is an average over a rolling seven-day window, with spend able to run up to twice your number on a given day. Minimums are set per billing currency — 725 INR, 25 USD, 15 GBP, 15 EUR, 25 AUD.

Bid adjustments exist in exactly one form: an audience bid multiplier between 0.1x and 10x at ad group level, applied when a viewer matches one of your custom audiences. Highest matching multiplier wins. No device, location, schedule or demographic modifiers.

Audiences: thinner than either platform

Meta gives you pixel-built custom audiences, lookalikes, and a deep interest and behaviour taxonomy. Google gives you Customer Match, remarketing lists, in-market and affinity segments, and similar audiences. ChatGPT gives you one thing: lists you upload yourself.

Custom audiences are built from emails, phone numbers, SHA-256 hashed emails or phones, or Google Advertising IDs. You can include them at campaign level, exclude them at campaign level, or attach a bid multiplier at ad group level.

The catch is size. Inclusion targeting and bid adjustments require at least 25,000 matched users, with 100,000 recommended. Below 25,000, a list can only be used for exclusions. Both Google and Meta let you activate far smaller lists, so this threshold quietly disqualifies most SMB customer databases from anything except suppression.

There are no lookalikes, no interest segments, and no pixel-built retargeting audiences. You also get no demographic targeting at all — no age, gender, income or parental status — and no demographic data comes back in reporting.

Creative: Meta’s playbook in Google’s format

The ad unit is six components: title, copy, image, landing page URL, advertiser name and logo. Text-first, closer to a responsive search ad than a Meta creative — but with no headline or description permutations, and no extensions of any kind. No sitelinks, callouts, structured snippets or call extensions.

Where it flips to Meta thinking is volume. OpenAI’s guidance is to run a high number of genuinely different creatives per offer rather than three near-identical variants, because every distinct angle gives the matching system another way to find you a relevant conversation. That’s the Meta creative-testing rhythm, not the Google one.

Two operational notes that catch people out. You can append your own UTM parameters to destination URLs, so GA4 and CRM attribution work normally. And your landing page must not block OpenAI’s crawlers, OAI-AdsBot and OAI-SearchBot — if the client runs an aggressive firewall or bot-blocking plugin, check this before launch, because it fails silently.

Reporting: the difference that actually changes your job

You get impressions, clicks, spend, CTR, average CPC, average CPM, and conversions where measurement is configured. Table view by campaign, ad group and ad; trend charts; CSV export for daily or cumulative values. Bulk upload and bulk edits both exist, which matters if you’re launching more than a handful of ad groups.

What you don’t get is the thing search marketers live in: there is no search terms report. No query data, no conversation data, nothing that tells you what triggered an impression.

That single absence rewires the optimisation loop. On Google you harvest search terms, add negatives, split out winners and refine match types — a weekly cycle of tightening. On ChatGPT there’s nothing to harvest. You change hints, you change creative, and you read outcomes. That is Meta’s loop, not Google’s.

OpenAI also states plainly that it has no performance benchmarks yet by advertiser, industry or campaign type. Anyone quoting you a typical ChatGPT Ads CPC is guessing.

Can you use negative keywords on ChatGPT Ads?

No. There is no negative keyword field anywhere in Ads Manager — not at campaign level, not at ad group level. It doesn’t exist as a concept, because there are no keywords to negate. There’s no placement exclusion, no topic exclusion, and no brand-safety category blocklist either.

What you have instead, roughly in order of usefulness:

Custom audience exclusions. You can exclude people, not conversations. Upload existing customers or a suppression list and stop serving to them. Usefully, exclusion audiences carry no minimum size — unlike inclusion lists — so this works even for a small client list.

Tighter context hints. Narrowing what you describe narrows what you’re eligible for, indirectly. But since hints don’t guarantee or restrict delivery, treat this as influence rather than control.

Ad copy and landing page as filters. Both feed the relevance model. Copy that states plainly who the product is for pulls you out of mismatched contexts better than vague copy does, and it filters again at the click stage — which protects spend on CPC campaigns.

Campaign structure. Separate campaigns per product, market and intent theme, so a bad match in one doesn’t eat a shared budget.

For most advertisers this is an annoyance you manage with structure. For some it’s disqualifying — legal, medical, finance, gambling-adjacent, anything with competitor-conquesting rules, or any brand with a contractual adjacency requirement. If a compliance team needs a documented exclusion list, you can’t produce one here yet.

And because there’s no search terms report, you can’t even see what you’d want to negate. That’s the sharper problem: you’re missing the control and the diagnostic that would tell you whether you needed it. In the meantime, watch landing page bounce and time-on-site by campaign, and watch for a CPC campaign where clicks climb while conversion rate collapses. That’s your only real signal that you’re matching into the wrong conversations.

Everything that’s missing, in one list

Against Google Ads and Meta Ads both, ChatGPT Ads currently has no:

• Keywords or match types
• Negative keywords, placement exclusions or topic exclusions
• Search terms or query-level reporting
• Volume or forecasting tools
• Demographic targeting or demographic reporting
• Interest, in-market or behavioural segments
• Lookalike or similar audiences
• Pixel-built retargeting audiences
• Dayparting or ad scheduling
• Device bid adjustments
• Ad extensions
• Automated rules or scripts
• Published performance benchmarks

So what is it actually like?

The closest existing analogue isn’t Google Search or Meta feed — it’s Performance Max. Broad, opaque, creative-led, with the control surface deliberately reduced and the system asking you to trust it with the matching. The difference is that PMax sits on Google’s mature signal graph, while ChatGPT Ads is running one that’s a few months old.

Practically, that gives you a clear operating model: budget it like a Meta test, structure it like a search account, and optimise it like PMax. Creative iteration and hint refinement, not query mining.

Who should test it now

ChatGPT Ads suits advertisers with broad addressable markets, a clear and explainable value proposition, and tolerance for spending into a channel with no benchmarks. It works best when the offer is something people genuinely ask an assistant about — comparing tools, researching a purchase, solving a problem.

It suits you badly if you depend on negative keywords for brand safety, if your economics only work on tightly-matched bottom-funnel queries, or if you need to forecast before you commit.

The workable setup right now: a small ring-fenced test budget, one campaign per market, tightly-themed ad groups with genuinely descriptive context hints, a high volume of distinct creatives, conversion tracking live before launch, and UTM parameters on everything so your own analytics tell you what OpenAI’s reporting can’t.

Treat the first month as buying information rather than buying conversions. The platform is moving fast enough that half of this comparison will be out of date by Christmas — but the advertisers who learn how context hints behave now will be the ones who know what to do when the controls finally arrive.