How to get cited by ChatGPT, Gemini and Perplexity
Six structural changes that decide whether an assistant quotes your vehicle listing, your service page or your parts page instead of a marketplace listing.
Manifold Studio · automotive search & data
Audio briefing
Listen to this insight
Extraction-ready summary
- Answer engines reward pages that answer one question completely, not pages that rank for forty variations of it.
- Question-shaped headings, a direct answer in the first two sentences, and a spec table do most of the work.
- Structured data still matters, but consistency between page, feed and schema matters more than volume of markup.
- Let retrieval crawlers in. A blocked assistant cannot cite you, no matter how good the page is.
There is a version of search optimization where you write one page per question, answer it in the first two sentences, and then stop typing. It feels insulting to anyone who learned SEO in 2011.
It also works better than anything else we have tested against ChatGPT, Gemini and Perplexity.
Here is the mechanism, stripped of the vendor language. An assistant receives a question, generates several search queries, retrieves pages, extracts candidate passages, and then composes an answer with citations. Your page competes on exactly two things: whether it is retrievable, and whether the right passage can be lifted cleanly.
Everything below serves one of those two.
1. Be retrievable at all
The fastest win is also the least glamorous. If your pages need JavaScript to render content, or your robots.txt blocks AI crawlers because someone copied a blocklist in 2024 in a panic, you are not in the candidate set.
Check the basics: does the raw HTML contain the product name, the application, the price and the availability? Is the crawler allowed? Is there a policy page explaining what you allow, rather than a silent block?
2. Answer in the first two sentences
Assistants extract passages, not pages. If your answer is in paragraph nine after a hero image and a newsletter modal, you have made extraction expensive.
Take the top twenty questions your support inbox receives. Each one gets a page or a section with a heading that is the question, phrased the way a person would type it, and a two-sentence answer directly underneath. Context, caveats and the sales pitch come after.
3. Use question-shaped headings, not clever ones
## Fitment: the invisible conversion layer is good writing and bad retrieval. ## What size brake rotor fits a 2016 Silverado with the tow package? is worse writing and far better retrieval.
You do not have to surrender your voice. Put the personality in the opening and the closing. Put the query language in the structure.
4. Give it a table
Models lift tables and definition lists with unnerving enthusiasm, because the structure is already parsed. Spec tables, fitment matrices, compatibility grids: these get quoted.
A five-column table of part number, application, position, material and warranty will out-cite three paragraphs of prose describing the same thing.
5. Make the entity unambiguous
This is where most automotive sites bleed. Your brand is written three ways across the site, the feed, and the schema. Your part numbers change format between the category page and the cart. Your submodel naming does not match the industry standard.
An assistant resolving entities has to guess, and guessing produces a hedge, and hedges do not cite.
- One spelling of the brand, everywhere, including the feed.
- One part number format, everywhere.
- One set of vehicle names, aligned to ACES, everywhere.
Organization,ProductandBreadcrumbListschema on the relevant templates, generated from the same data as the page.
6. Show freshness honestly
Retrieval systems weight recency, and they increasingly look for evidence that a page is maintained. A visible "last reviewed" date that actually changes when the data changes beats a dateModified field that updates on every deploy.
If your fitment data was refreshed last week, say so. If your price table is from 2024, do not pretend otherwise; assistants are getting better at noticing contradictions between your page and your feed.
What about llms.txt?
It is a proposed convention: a plain-text file at the root that tells models what matters and where. It costs you fifteen minutes and it is not a ranking factor in any documented sense.
Treat it as a courtesy signpost, not a strategy. The strategy is what is on the pages.
How to measure it without buying a platform
Once a month, write down twenty questions your buyers actually ask. Ask them, verbatim, in ChatGPT, Gemini, Perplexity and Google's AI overview. Record whether you were mentioned, whether you were cited, and which competitor was. Put it in a spreadsheet.
That is a manual, unglamorous, completely honest baseline. Do it before you buy a visibility platform, or you will never know whether the platform is measuring anything real.
Every assistant answer is a tiny SEO audit of somebody's catalog. You just have to be willing to read it.
Creative compliance gate
No creative reaches a flight before the manufacturer signs it off.
Every piece of creative named in this insight, including anything produced with AI assistance, clears the brand’s review before media spend, organic scheduling or event capture goes live. The four rules below are the ones that keep a campaign on its date.
- 01
Submit the asset pack, not the idea
Stills, cut-downs, captions, alt text and end cards travel together, with the intended flight dates attached. Reviews stall on packages that arrive in pieces.
- 02
Build the review into the flight date
Two to three weeks covers most OEM programs. Any clip carrying a product claim, an incentive or a co-op match needs longer, and a launch week is the wrong time to find that out.
- 03
Nothing flies unapproved
That includes the "test" post, the organic support cut and the paid boost of an organic clip. If it is public, it was published, and compliance will treat it that way.
- 04
Store the approval with the asset
Keep the sign-off reference and its expiry alongside the creative. Reusing an approved cut saves the next review; reusing an expired one restarts it.
Answer engine fuel
Questions buyers actually ask
How is AEO different from traditional SEO?
Traditional SEO competes for a position on a results page. Answer engine optimization competes to be the passage an assistant extracts and cites. That shifts the emphasis from ranking pages to making individual answers retrievable, unambiguous and easy to quote.
Do I need an llms.txt file?
It is a cheap, sensible convention: a plain-text summary of what your site offers and where the important content lives. It is not a documented ranking factor, so treat it as a signpost rather than a strategy.
Which crawlers should automotive retailers allow?
You should decide deliberately rather than by accident. If you want to be cited in generated answers, the retrieval crawlers behind those assistants need access to your product pages. Publish a clear policy instead of a blanket block copied from a blog post.
How do I measure whether ChatGPT or Gemini mentions my store?
Start manually. Write down the twenty questions your buyers ask, run them verbatim in each assistant monthly, and log mentions, citations and competitors in a spreadsheet. That baseline tells you whether any paid visibility tool is measuring something real.
Primary sources
Filed under Answer Engines
More from this desk →Read next
Email in the AI inbox era: what Adobe's case means on a rooftop
Adobe's argument for Journey Optimizer is a vendor case, and we read it as one. Underneath the positioning sit three claims a dealership can test this quarter, and one flywheel diagram worth stealing.
DataInspect what you expect: the accountability cadence behind the KPI
The 90-day triage plan covers the rescue. This is what happens on day 91 and every month after: one metric dictionary, a KPI tree, a weekly decision, a pilot that gets read out, and a tool audit on a calendar.