How to rank in ChatGPT

ChatGPT can put a brand in front of a buyer in three different places, and three different systems feed them. Winning one does nothing for the other two.

The three surfaces at a glance

• Written answers with citations. Fed by live retrieval during the conversation plus the model's training memory. Won by being crawlable, quotable and widely referenced elsewhere.
• Product results and Instant Checkout. Fed by merchant product data and the Agentic Commerce Protocol, the open standard from OpenAI and Stripe that carries the cart and payment. Won by feed quality and checkout integration.
• Sponsored placements. Sold by OpenAI itself through its own Ads Manager at ads.openai.com, with US testing that began on 9 February 2026 as a phased rollout. Advertisers already buying through Amazon DSP can extend campaigns into the same surface under an Amazon Ads pilot opened on 10 September 2026. Won with budget and targeting rather than content.

Two mechanisms sit behind the first surface and they behave differently. Training memory is what the model already absorbed about your brand from the open web, slow to change and impossible to edit directly, while live browsing is what it fetches mid-conversation and can shift within a crawl cycle.

One test separates them. Ask your buying question plainly, then ask the same question again with "search the web and cite your sources" appended, and compare the two answers.

The result tells you where the gap is. If you appear only in the second answer, retrieval works and the model does not know your brand. If you appear in neither, start with crawler access before touching any content.

The practical consequence is organizational. These are three workstreams with three owners: content and SEO own citations, ecommerce and ops own the feed and checkout, and paid media owns sponsored placement. Deeper detail on the commerce side sits on sell on ChatGPT and on the agentic commerce protocol.

Crawler access is the first thing to check, because every other change is wasted if the answer is no. OpenAI runs separate crawlers for separate jobs, and blocking each one costs you something different.

Which agent does what

• GPTBot. Collects content used for model training. Blocking it affects long-term memory, not live answers.
• OAI-SearchBot. Builds the search index behind answers that carry citations. Blocking it removes you from exactly the answers that link out.
• ChatGPT-User. Fetches a page during a conversation, for example when someone pastes your URL into the chat. Blocking it breaks the moment a prospect brings your site into the conversation themselves.

Most brands that are invisible in ChatGPT never made a decision about this. A blanket AI bot block got switched on at the CDN, or a robots.txt line was copied from a template, and nobody revisited it when the search crawler became a separate agent.

How to check in ten minutes

1. Read your robots.txt and list every AI user agent named, allowed or disallowed.
2. Check the CDN or WAF for managed bot rules, because those reject requests before robots.txt is ever read.
3. Search the last 30 days of server logs for each agent and count fetches per agent per day.
4. Confirm the pages they fetch return 200 with full content as plain HTML, with no JavaScript required.

Verify the current user agent names against OpenAI's published crawler documentation before editing anything, because that list changes and a stale Disallow line is worse than none.

One more constraint worth knowing: some large retailers deliberately block agent traffic. If most of your volume goes through marketplaces, the surface being read may not be your own site at all, which changes where the fix belongs.

Product results in ChatGPT are built from merchant data, not from your marketing pages. If the feed is thin or stale, no amount of content work will place you in an answer to "best waterproof backpack under $80".

Instant Checkout goes one step further than discovery. It uses the Agentic Commerce Protocol from OpenAI and Stripe to pass the cart and a payment token to the merchant, so the purchase can complete inside the conversation rather than on your site.

What a usable feed looks like

• One row per buyable variant. Parent-level rows lose the size, color or pack count the shopper actually asked for.
• Identifiers that resolve. GTIN or MPN where they exist, plus a canonical product URL that returns 200 with no redirect chain.
• Price and availability matching the live page. A nightly file is not real time, and a price mismatch at checkout loses the sale and the placement together.
• Literal attributes. "Machine washable, 100% merino, 35 litre" is matchable. "Built for the everyday adventurer" is not.
• Shipping and returns as structured fields. Agents filter on them the way a shopper filters a marketplace.

OpenAI's product feed specification requires less than you would expect and rewards more. At the product level only an id and a set of variants are required, and each variant needs an id and a title. Everything that decides whether you get matched, price, availability, categories, media and seller, sits in the optional fields, so treat the required list as the floor and not the target.

The known weak spot is complex variants. Agents still struggle when size, pack count and style have to be resolved together, so a catalog that flattens those into one option string gets matched to the wrong item or skipped entirely.

A quick way to test this yourself: ask for your own product with three constraints at once, such as color, size and quantity. If the assistant returns the right brand and the wrong variant, the problem is your variant structure, not your visibility.

Paid placement inside ChatGPT is now real, and there are two doors into the same surface. OpenAI sells it directly, bought through its own Ads Manager at ads.openai.com, with US testing that began on 9 February 2026 as a deliberate, phased rollout. On 10 September 2026 Amazon Ads opened the second door, a pilot letting select US advertisers extend campaigns into ChatGPT as conversational ads bought through Amazon DSP, with Delta Vacations among the first testers.

Three things follow for a merchant. The rollout is phased and availability varies by country, so the first question is whether you can buy where you sell. Buying happens in OpenAI's own Ads Manager, or through Amazon DSP for advertisers already buying there, and the unit is a sponsored product card below the response, labeled as sponsored and set apart from the answer.

What ads do not do is equally important. They do not make the model cite your page inside a written answer, and they do not repair a broken feed. A sponsored slot below an answer that recommends three competitors is an expensive way to look like the outsider.

Before you put budget behind it

• Confirm availability first. Availability varies by country and keeps expanding, so check OpenAI's published Ads Manager availability page before planning a quarter around it.
• Run the organic diagnostic first. Knowing whether you are cited, merely mentioned or absent changes what the ad has to carry.
• Keep the feed in sync. Paid traffic landing on a stale price or an out of stock variant wastes the spend twice.
• Report it as paid media. It is bought and measured in OpenAI's own Ads Manager or in Amazon DSP, not in your SEO reporting.

Treat the three surfaces as a sequence rather than a choice. Crawler access and quotable content make you a candidate, the feed makes you buyable, and ads buy reach on top of a position you already hold.

Keep the measurement separate as well. Ad reporting comes from the ad platform, citation share comes from your prompt panel, and product surface presence comes from buying-intent prompts, so a single blended dashboard will hide which lever moved.

There is no rank to look up, so you sample. A fixed set of prompts, run the same way each month, is the closest thing to position data that exists.

1. Write 25 prompts in buyer language across four intents: discovery ("best running shoes for flat feet"), comparison ("brand A vs brand B"), objection ("is brand A worth the price"), and purchase ("buy a 35 litre waterproof backpack under $80").
2. Freeze the wording in a spreadsheet. Any rewording later invalidates the trend line you are building.
3. Run each prompt three times in a temporary chat while logged out, then once more with "search the web and cite your sources" appended.
4. Record per run: brand mentioned, page cited with the exact URL, position in any list, whether product results or a buy option appeared, and which competitors showed up.
5. Rerun the purchase prompts on their own to test the shopping surface, and note whether your products appear at all.
6. Open the competitor URLs that got cited and list what those pages have that yours do not, which is usually a sourced number, a comparison list and a plain answer paragraph near the top.
7. Check crawler access and feed status as described above, then date-stamp the whole sheet.
8. Repeat monthly on the same day, same wording, same order.

Reading the result

• Absent everywhere, no crawler hits in logs. Access problem. Fix robots and CDN rules first and wait a crawl cycle.
• Named only when you force a web search. Retrieval works, training memory does not know you, so the work is third-party coverage.
• Named in discovery, absent in purchase prompts. Feed problem rather than a content problem.
• Cited but always last in the list. Comparison problem, usually a missing attribute breakdown or hidden pricing.

Four changes take days rather than quarters. Each is written as what to change and what good looks like.

1. Separate the crawler decisions

What to change: allow the search crawler and the in-conversation fetcher, and treat the training crawler as its own licensing question. What good looks like: server logs show daily fetches from the search crawler, and no CDN rule silently overrides robots.txt.

2. Put an answer block on every commercial page

What to change: add a 40 to 60 word answer directly under the H1 that states what the product is, who it is for and what it costs. What good looks like: that block still makes sense pasted alone into a blank document.

3. Publish prices and comparisons as text

What to change: numbers locked in images, PDFs or JavaScript widgets get rebuilt as plain HTML, including the comparison points buyers actually ask about. What good looks like: curl on the page returns every figure a shopper would weigh.

4. Fix the five feed fields that break matching

What to change: variant-level rows, price sync, availability, canonical URLs and literal attribute names. What good looks like: a spot check of ten SKUs shows feed price and stock matching the live page within minutes, not hours.

One extra pass is worth the time. Make sure your product facts read identically on your site, your marketplace listings and your social profiles, because a stale third-party listing can outvote your own page when a model reconciles conflicting numbers. Ten minutes on the three biggest listings usually finds at least one price or spec that no longer matches.

The quick fixes make you eligible. Becoming the default recommendation takes longer, and most of it sits outside your content team.

Real-time data instead of scheduled files

Shopify's guidance on agentic-ready product data sets the bar as structured, machine-parsable and real-time, with variant grouping, precise taxonomy, literal descriptions and live price and inventory. Shopify reports AI-referred orders grew nearly 13x year over year and that AI-referred visitors convert at nearly 50% higher rates than organic search, documented in its agentic-ready product data guide.

Rebuild variant architecture

Flattening size, pack count and style into one string is the single most common reason a correct brand match returns the wrong product. Restructuring that touches your PIM, your theme and your feed, and it usually takes a quarter, which is why teams keep postponing it.

Earn the third-party coverage models lean on

Assistants quote roundups, forums, review sites, retailer pages and video transcripts far more often than vendor sites. That is partnership and PR work with a two to three quarter horizon, and it is what fixes an absence that crawler settings cannot explain.

Decide about in-chat checkout

Instant Checkout moves the purchase out of your funnel, which affects returns handling, support, attribution and how you collect customer data. Work through those with ops and finance before engineering, because the integration is the easy half.

Sequence matters more than speed here. Feed accuracy first, then variant structure, then third-party coverage, with the checkout decision taken whenever your operations can actually support it.

Keep the monthly prompt panel running the whole time. It is the only way to attribute a change in visibility to one of these four tracks rather than to a model update you had no control over.

QuickAds does not run your robots.txt or build your checkout integration. What we do sits in two places: the product data that shopping surfaces read, and the creative that has to exist once you are visible.

On the data side we fix Shopify catalog structure: variant grouping, taxonomy, literal product fields and live price and stock sync. That is the same work that decides whether a size plus pack count plus style prompt returns the right SKU, and it is usually the fastest fix available to a merchant.

On the creative side we produce 100+ creatives per month on a 5-7 day turnaround, with creative intelligence trained on 32M+ ads behind the concepting. That matters for the paid surface, since sponsored placement is bought with creative and budget, and for everything else you run alongside it.

We also generate catalog video per SKU with structured metadata and schema attached, so the video is parsable by an agent rather than an opaque file. Untagged video contributes nothing to retrieval, and most brands have libraries sitting in that state.

The coverage point: we run intelligence, strategy, production and campaign management as one chain, while analytics-only tools cover measurement and premium production shops cover production. The handoffs between point solutions are where the weeks disappear. The production side is set out on creative as a service, and the engine-agnostic version of this playbook sits on LLM SEO.

Software starts at $299 per month and managed engagements start at $5,000 per month, so the catalog and video work can be tested on a subset of SKUs before it becomes a program.