Why per-platform matters
Each AI engine has its own crawler, its own index, and its own citation patterns. Tactics that get you cited in ChatGPT will not necessarily land you in Perplexity, and vice versa. The most common mistake we see at NOVA is brands writing AI search content as if all AI engines worked the same. They do not.
This post is the tactical companion to What is GEO, which defines the discipline, and to AI search mechanics, which explains how the underlying systems pull content. If you want the why and the how-it-works, start there. This post is the what-to-actually-do.
What 403,000 prompts reveal about who AI recommends
Before the platform-by-platform tactics, it helps to know what actually correlates with being recommended. Ben Wills recently published the largest study of AI visibility factors I have seen so far: 403,000 prompts run across 10 different AI models, covering 100 industries including many local service categories.
The strongest signals that correlated with a business being recommended, in the professional-services verticals the study highlighted, were:
- Appearing in Google's search results. Anywhere on page 1 was enough. This was the single strongest correlation in the study, which means classical SEO is still the foundation under every AI engine.
- A homepage that clearly states what you do and where. Models lean heavily on how plainly your homepage describes your services and location.
- Backlinks and domain authority. The old currency still spends.
- A presence on Wikidata. Far easier to get than a Wikipedia page, and it feeds the knowledge graphs several engines draw from.
- Mentions in relevant Reddit discussions. Reddit's licensing deals mean its threads are baked into more than one model's training and retrieval.
If you only take one thing from this post, take this: AI visibility is not a separate discipline you bolt on after SEO. Ranking on page 1, saying plainly what you do, and building real authority is most of the work. The platform-specific tactics below are the tuning on top.
ChatGPT
ChatGPT cites web sources in two situations: when the user has web browsing enabled, and when the model has training data familiarity with your brand. Each path requires different work.
How ChatGPT decides what to cite
When web browsing is active, ChatGPT runs a search behind the scenes, fetches a small number of top results, and synthesizes an answer. Bing's index has historically been the gatekeeper: if you are not in Bing, you are largely invisible to ChatGPT's browsing mode.
But it is not just one search lane anymore. Mark Williams-Cook recently dug into ChatGPT's search behavior and found a hidden label on each result called result_source, with three values:
- serp: normal web search results, the lane you influence with classical SEO and Bing indexing.
- labrador: trusted or licensed sources such as Reuters, Wikipedia, and product review platforms. You influence this lane through review-site presence and entity building, not through your own site.
- bright: what appears to be structured data feeds, likely from Bright Data. You do not control this lane directly, but it rewards having consistent, machine-readable business data everywhere.
The practical takeaway: your own website only competes in one of the three lanes. The review-platform and entity work that used to feel optional is how you show up in the second, and consistent structured data everywhere is your ticket to the third.
For training data familiarity, ChatGPT pulls from sources baked into the model. This favors high-traffic websites, news mentions, Wikipedia, and review platforms over time.
What to do
Get into Bing's index. Submit your sitemap to Bing Webmaster Tools. Most small businesses skip this and lose ChatGPT visibility by default.
Allow GPTBot in robots.txt. The default OpenAI crawler is GPTBot. If your robots.txt blocks it, ChatGPT cannot read fresh content from your site. Confirm access.
Structure for direct answers. ChatGPT favors content where the first one or two sentences under each H2 directly answer the heading. Long preambles get skipped.
Use comparison tables. ChatGPT extracts HTML tables almost verbatim. Any pricing comparison, feature comparison, or before-after framing belongs in a table.
Build review platform presence. Trustpilot, G2, Capterra, and Clutch are weighted heavily in ChatGPT's authority signals. Brands with active profiles get cited around 3x as often as those without.
Perplexity
Perplexity does not use Bing's or Google's index. It runs its own crawler (PerplexityBot) and maintains its own index optimized for answer generation. That changes the work.
How Perplexity decides what to cite
Perplexity reads its own crawl, weighted by content freshness, structured data, and source authority. It tends to favor sources that publish dated, well-cited content with visible authorship. It also weights recency more heavily than ChatGPT does.
What to do
Allow PerplexityBot in robots.txt. Confirm access. Many sites block it accidentally through a generic AI bot block.
Add visible publish dates and last-updated dates. Perplexity weights freshness more than other engines. A page dated 2026 outperforms an undated page with identical content.
Add JSON-LD structured data. Perplexity favors machine-readable content. Article, FAQPage, and HowTo schema all help.
Cite your own sources. Perplexity favors sources that themselves cite. Inline references with links inside your content increase your odds of being read as authoritative.
Set up llms.txt. llms.txt is an emerging standard that helps AI engines, including Perplexity, understand your site's purpose at a glance.
Claude
Claude uses Anthropic's web search when available and pulls from training data otherwise. Citation behavior is more conservative than ChatGPT or Perplexity. Claude tends to cite fewer sources but cites them more deliberately.
How Claude decides what to cite
Claude favors sources with clear authorship, transparent reasoning, and high editorial quality. It is more sensitive to content that reads as marketing copy and tends to skip it. Long-form, well-reasoned content with evidence tends to get cited.
What to do
Allow ClaudeBot in robots.txt. The Anthropic crawler is ClaudeBot.
Lead with evidence, not claims. Marketing assertions like best in class or leading provider get filtered out of Claude's reasoning. Specific data, examples, and case studies do not.
Show your work. If you make a claim about how something works, link to or describe the source. Claude tends to favor content that exhibits its own reasoning.
Avoid AI-generated filler. Claude is unusually good at detecting low-quality AI-generated copy. Pages that read like generic AI content get demoted.
Google AI Overviews
AI Overviews appear at the top of about 30% of US Google searches. They are powered by Gemini and pull from Google's index, layered with AI synthesis.
How AI Overviews decide what to cite
AI Overviews use Google's existing ranking signals (links, content quality, E-E-A-T) plus AI-specific factors. The same SEO work that earns you a top-ten ranking also earns you AI Overview citations, but with extra weight on direct answers, structured data, and content that maps cleanly to question-answer pairs.
What to do
Stay in Google's index. Same as classical SEO. If you are not in the top ten for the query, you are unlikely to be cited. Fix indexing first.
Strengthen E-E-A-T signals. Visible author bios with credentials, publish dates, last-updated dates, and citation of authoritative sources all help.
Keep the FAQ content, skip the FAQPage rich-result chase. Google deprecated FAQ rich results for most sites, so FAQPage markup no longer earns the visual treatment it used to. The question-and-answer format itself still matters a great deal, because AI Overviews map H2s and question-answer pairs to sub-queries. Write the FAQ content, mark the page up with Article schema, and do not expect rich results from FAQPage markup.
Write the answer in the first sentence. AI Overviews extract the opening line under each heading. Lead with the answer.
Match the search query in your H2s. AI Overviews map H2s to sub-queries during generation. A page with H2s like How much does this cost? outperforms a page where the same content lives under Pricing details.
Gemini
Gemini sits inside Google's products (Search, Android, Workspace) and uses Google's index. It overlaps with AI Overviews more than people realize, and the optimization work is largely shared.
What to do
The same work that earns AI Overview visibility tends to earn Gemini visibility. There are two additional considerations.
Consider Google Workspace context. Gemini inside Workspace can pull from Google Docs and other corporate context for some queries. This affects B2B more than B2C.
Monitor Gemini directly. Gemini's citations are often slightly different from AI Overviews even on the same query. Test both.
The local layer: Google's AI Local 6-Pack
If you serve local customers, there is a change happening right now that deserves its own section. Joy Hawkins recently spotted Google testing AI-driven local results, which Darren Shaw has taken to calling the "AI Local 6-Pack": instead of the familiar map pack, Google assembles an AI-generated set of local recommendations.
Everything in this post still applies, and two Google Business Profile signals get heavier:
Your categories are feeding Google's AI. Your primary category remains one of the most important fields on your profile, and you can add up to nine additional categories. As AI queries get more specific ("best quiet restaurant with vegetarian options near me"), accurate categorization is how Google's AI knows you belong in the answer. Cover every legitimate service you offer, and nothing you do not.
Your reviews are being mined for details. Whitespark analyzed 765 Google Business Profiles and found Google may ask reviewers up to 14 category-specific follow-up questions when they leave a review: how much did it cost, how was the parking, did they have vegetarian options. Those answers exist so Google's AI knows exactly when to recommend you. The move: rotate your review requests. Ask one batch of customers to mention service and atmosphere, the next batch to mention parking and pricing, so that over time your reviews cover the attributes AI queries filter on.
And the same playbook applies for the 6-Pack itself: keep doing SEO as usual, make sure your content answers the questions people actually ask in AI chats, and build your brand where the models are listening, meaning Reddit, YouTube, and active social profiles, not just your own site.
Common ground across all platforms
Across every platform above, a few things consistently help.
| Tactic | Why it matters | Effort |
|---|---|---|
| Structured data (JSON-LD schema) | Machine-readable content gets read more reliably | Low |
| Visible author bios + dates | E-E-A-T signal across all engines | Low |
| Direct-answer first paragraph under each H2 | All engines extract opening sentences | Low |
| Comparison content as tables | All engines extract tables verbatim | Medium |
| Review platform presence (Trustpilot, G2, Clutch) | Authority signal across all engines | Medium |
| Allow AI bots in robots.txt | Prevents accidental invisibility | Low |
| Set up llms.txt | Emerging standard, low effort, high signal | Low |
| Wikidata entity listing | Feeds knowledge graphs multiple engines draw from; correlated with AI recommendations in the 403,000-prompt study | Low |
If your team can only commit to a few of these, start with the low-effort ones at the top of the table.
How to test your visibility
Run the same query in five places once a month, and track which platforms cite you. Use a query your customers actually ask, not a query you wish they asked.
A simple monthly tracker:
| Query | ChatGPT | Perplexity | Claude | AI Overview | Gemini |
|---|---|---|---|---|---|
| [Your service] near me | Cited / Not | Cited / Not | Cited / Not | Cited / Not | Cited / Not |
| Best [your category] in [city] | |||||
| How much does [your service] cost |
When a competitor is cited and you are not, fetch the page they got cited from and compare it to yours. The differences usually fall into the buckets in the Common ground table above.
For a free assessment of where your business stands across AI engines, take the AI Readiness Scorecard.
FAQ
Why am I cited in ChatGPT but not Perplexity?
Most likely Bing has indexed your site but Perplexity's crawler has not, or you blocked PerplexityBot in robots.txt. Confirm access first.
Do I need to do this work for every AI engine separately?
The foundational work overlaps. About 70% of the tactics serve all engines. The remaining 30% is platform-specific tuning. Start with the common ground.
Does ranking in Google still matter if my customers use ChatGPT?
Yes, more than anything else. The largest study of AI visibility to date (403,000 prompts across 10 models) found that appearing anywhere on page 1 of Google was the strongest single correlate of being recommended by AI. The engines disagree on many things, but they all trust search results as a quality filter. If you are not ranking, fix that first.
Should I block AI crawlers if I do not want to be in AI search?
You can, but you will not show up in AI answers, and your visibility loss compounds as adoption spreads. We recommend allowing crawlers and instead setting up llms.txt to control how your site is described.
How long until I see results from AI optimization?
Faster than traditional SEO, in our experience. Some changes (robots.txt, schema, llms.txt) can take effect within days. Authority and citation patterns take longer, usually 60-90 days.
Is this work worth the effort for a small Boston business?
Almost always yes. AI engines name two or three businesses per recommendation query. Being one of those three is a substantial advantage that is much cheaper to capture now, before competitors notice.
