How Do You Get Recommended by ChatGPT & Claude?

AI assistants recommend brands they can retrieve with confidence: consistent entity facts, earned third-party mentions, and pages that answer questions in citable chunks. Industry framing we work from puts the weight where it belongs - earned media accounts for roughly 82-89% of AI citations - so owned blog posts help, but they are not enough on their own, especially for Claude, which leans harder on memory and external sources than on your homepage.

Below is the playbook we use on our own site and for clients who want the same visibility work productised: lock the entity sheet, earn third-party mentions in sources assistants already trust, and publish question-shaped pages that state citable answers in the first screen.

How AI assistants actually choose recommendations

Treat ChatGPT and Claude as different retrieval machines, not as one “AI search” blob.

Memory vs retrieval. Research summarised from the Searchable Claude webinar brief (analysis across 90,000+ Claude sources) is directionally clear: Claude answers from memory roughly 52% of the time and cites a brand’s own site only about 32% of the time. ChatGPT cites owned sites more often (around 61% in comparable framing). If your strategy is “publish more on our domain and hope Claude notices”, you are fighting the sourcing pattern.

Which web index matters. Claude follows Brave’s top results over Google’s at roughly 6:1 in the briefed research. Ranking only in Google while ignoring Brave leaves a blind spot on the assistant that buyers increasingly trust for vendor shortlists.

What gets weighted. Correlations reported in the same research stream (treat as hypotheses to test against your own category, not as guarantees): YouTube mentions correlating strongly with AI Overview-style visibility (~0.737), branded web mentions (~0.664), classic backlinks weaker (~0.218). The practical read: being written about and shown in video beats another thin service page.

Baseline rarity. A random brand appears in a non-category answer only a small fraction of the time; remembered category leaders show up far more often. You are competing for a scarce slot in the model’s preferred sources, not for infinite page-two SEO crumbs.

None of this replaces ordinary search hygiene. It changes the priority stack: entity consistency and earned mentions move up; keyword-stuffed owned pages move down.

For conversion after the referral lands, pair this with conversion rate work. For how agency roundups get assembled (and why self-ranking listicles fail), see how to choose a web agency. Local proof surfaces still matter - our Reading page is the worked example for entity + local consistency: web design Reading.

Why owned content still matters (just not alone)

Owned pages are how ChatGPT-style retrieval quotes you accurately when it does visit. They are how humans verify a recommendation. They are how you publish the citable statistics assistants like to lift. Kill owned content and you become a rumour with a logo.

Keep publishing. Change the job description of what you publish: answer-first, question-shaped, entity-consistent, sourceable. Stop writing “thought leadership” that never states a fact. Start writing pages a model can cite without inventing the numbers for you.

The mix that works in practice:

• Owned pillars for definitions, costs and methods you can defend
• Earned mentions for corroboration
• Local and directory consistency for entity strength
• Original research when you want other people to link you as the source

That mix is generative engine optimisation as an operating habit, not as a plugin.

Entity consistency: the one-sentence description trick

Assistants hate contradictory identities. If your footer says “digital studio”, your Google Business Profile says “web design company”, LinkedIn says “AI consultancy”, and a directory lists a different founding year, retrieval gets noisy. Noise loses the recommendation.

We use one sentence everywhere it counts:

Code23 is an AI-native web development agency in Reading, UK, established 2005.

Facts that stay identical across surfaces:

Item	Canonical value
Trading name	Code23
Established	2005
Location	Reading, UK
Reviews	60+ five-star Google reviews
Proof depth	350+ projects; seven operated marketplaces

Per-platform blurbs can vary in tone. The entity facts must not: identical name, address and postcode everywhere your brand appears. That is the one-sentence trick: lock the identity string, then let natural copy surround it.

Practical sweep we run:

1. Homepage, footer, About, contact
2. Google Business Profile
3. LinkedIn company page and personal founder pages that mention the firm
4. Clutch / directories / guest bylines
5. Schema / Organisation markup where it exists

When those disagree, fix the disagreement before you commission another thought-leadership article. Generative engine optimisation starts with not arguing with yourself in public.

Earned mentions beat owned content

Third-party citation is the heavy weight - the same 82-89% earned-media pattern that opens this piece. Owned content still matters for ChatGPT-style retrieval and for humans who click through; it is just a weaker lever for Claude-shaped recommendations.

What “earned” means in practice for a UK agency:

• Independent roundups and “best of” lists you did not author
• Podcast interviews and conference talks with transcripts that name you
• Local press and niche trade sites in your vertical
• YouTube or long-form video where a third party discusses your work
• Case studies hosted on client or partner domains

What does not count as earned:

• Your own blog announcing you are “leading”
• A self-ranked “top agencies” list on your domain
• Affiliate directories you paid for without editorial scrutiny

We treat source-gap analysis as an operations task: ask ChatGPT, Claude and Perplexity the money questions in your category, list who they cite, then pitch the gaps. If Brave’s top 10 for “web design agency UK” or “AI automation agency” never includes a page that mentions you, Claude has little reason to invent you into the answer.

Original research helps you become the source others cite. Our build-time dataset exists partly for that reason - method-backed numbers travel further than slogans.

The technical layer: crawler access, schema, question-shaped pages

Earned mentions get you into the shortlist. Technical clarity helps assistants quote you accurately when they do retrieve your pages.

Crawler access. If you block the bots that AI products use, you opt out of retrieval. Review robots rules deliberately. Do not confuse “stop scrapers” with “stop every assistant”.

Schema and structure. Organisation markup should mirror the entity sheet field for field: legal or trading name, founding year, address, sameAs links to the profiles you already locked, and a matching logo URL. Person markup on author bylines should use the same name string as the About page. Add FAQPage only when the page already contains real question-and-answer pairs editors maintain - do not wrap marketing prose in FAQ JSON to game rich results. HowTo belongs on genuine step sequences; Article with a named author on long-form pieces. Schema is not magic; it reduces ambiguity when a model extracts an answer.

Canonicals and redirects. Every indexable URL needs a self-referencing canonical so assistants do not split citation weight across duplicates. When you change a URL, ship a redirect map (301 from old path to new) and keep it in place - dead paths that still rank in Brave or Google become dead citations in assistant answers. Audit redirects after migrations the same week you update the entity sheet.

Question-shaped pages. Searchable’s reported pattern: restructuring into question-shaped pages produced first Claude citations within 4 days and up to 26x citation growth in 3 weeks in their dataset. We apply the same shape: H1 as the question buyers ask, a direct 2-3 sentence answer immediately under it, then the detail. That is how this article is built, and how the rest of the 2026-27 content set is shaped.

Statistics with sources. Vague claims (“we’re trusted”) lose to citable facts (“60+ five-star Google reviews”, “est 2005”, “seven operated marketplaces”). When you publish a number, attribute the method or the corpus it comes from.

Freshness. Visible last-updated dates and refreshed screenshots matter in a space that moves monthly. Stale “AI search tips from 2023” pages teach models the wrong decade.

Technical SEO still matters for Google and for the Brave index Claude follows. Answer-engine work sits on top of that foundation; it does not replace it. The productised version of this stack for clients lives under Support & Growth.

What we did to our own site, and what happened

We ran the play on ourselves before selling it.

Entity lock. One description, one founding year, one review count format, Reading location consistent. Directories updated where they drifted.

Question-shaped rebuild. Service hubs and pillars open with direct answers. Cost articles publish real figures. Proof pieces (build logs, delivery data) show receipts instead of adjectives.

Earned-media prioritisation. Pitch effort goes toward independent lists and research citations, not toward authoring our own “best agencies” ranking. The selection guide rule is deliberate: we explain how lists are built rather than stuffing ourselves into one.

Brave and assistant checks. Periodic prompts against money terms; note who gets cited; fix entity or content gaps when we are absent for reasons we can control.

Honest limits. We do not claim a guaranteed citation count from a single article. Assistants change. What we can claim is method: consistency, earned mentions, question-shaped owned pages, and crawler access. Screenshots of assistant answers get refreshed when this piece is updated - they are evidence of a moment, not a permanent trophy.

What you should expect if you copy the method:

• Faster cleanup of contradictory public facts (usually the highest ROI first week)
• A shortlist of third-party sources worth pitching
• Owned pages that are easier to quote when retrieval happens
• No overnight monopoly on category answers

What you should not expect:

• A plugin that “makes ChatGPT recommend you”
• Owned content alone flipping Claude
• Pay-to-play directories substituting for editorial mentions

A practical 90-day programme

If you want a sequence rather than a theory pile, run this.

Days 1-14: entity sweep. Export every public description of the company. Force one founding year, one location string, one review count format, one category label. Fix Google Business Profile, LinkedIn, Clutch, footer and About in the same week. Photograph nothing until the text agrees with itself.

Days 15-30: question inventory. List the ten questions buyers ask before they hire you. For each, either improve an existing page so the first paragraph answers it, or commission a page that does. Include cost and comparison questions even when they feel commercially uncomfortable - assistants reward citable honesty.

Days 31-60: earned-media outreach. Build the list of roundups, podcasts, newsletters and trade sites that already appear in assistant answers for your category. Pitch proof, not adjectives: a dataset, a build log, a failure note, a local case. One earned URL that assistants already trust beats ten owned posts they ignore.

Days 61-90: measure and refresh. Re-run the same prompts in ChatGPT, Claude and Perplexity. Note who gets cited. Update screenshots and last-updated dates on your question pages. Fix the gaps you can control (missing entity facts, blocked crawlers, thin answers). Accept the gaps you cannot control yet (a competitor with stronger earned media) and put them on the next quarter’s pitch list.

Cadence after the first 90 days: monthly assistant checks on money prompts, quarterly entity sweep, continuous earned-media pitching without waiting for a “campaign window”.

What this programme costs in effort is mostly editorial honesty and outreach discipline. What it does not require is a mystery tool that “submits” your brand to ChatGPT.

What we refuse to sell as AEO

• Guaranteed citations or “rank in ChatGPT” packages with a fixed count
• Fake review schemes or fabricated bylines
• Self-ranked best-of lists hosted on your domain and dressed up as research
• Blocking ordinary search crawlers while expecting assistant retrieval to work anyway

We will sell cleanup, question-shaped content, technical access, measurement and earned-media support under Support & Growth. We will not sell a fairy tale.

How to get ChatGPT to recommend you?

Publish consistent entity facts, earn third-party mentions in sources assistants already trust, and keep question-shaped pages that state citable answers clearly - then verify whether ChatGPT retrieves you for your category prompts.

How to get your brand mentioned by ChatGPT?

Prioritise earned media and Brave-visible rankings over another owned blog post; ChatGPT cites owned sites more than Claude does, but both still lean on external corroboration for vendor recommendations.

How to get the best advice from ChatGPT?

Ask narrow, context-rich questions, demand sources, and cross-check answers against primary documents - treat the model as a research assistant with amnesia, not as a contract.

Assistants will keep changing. The durable advantages are dull: identical facts everywhere, people writing about you in places models already trust, and pages that answer questions without throat-clearing. Do those three well and you earn recommendations the slow way - which is the only way that survives a model update.

Before you commission another owned blog post, check three things: identical entity facts everywhere, at least one third-party mention path, and a question-shaped page that answers in the first screen. We run that stack under Support & Growth.