AI Search Optimisation Checklist (Score /100)

Most businesses that never appear in ChatGPT do not have a content problem. They have a plumbing problem and a proof problem: a site the crawlers cannot read, two versions of the brand, experts with no names, and almost nobody else saying they are good. This playbook gives you the audit we use, with the scoring rubric, the prompt test, a log template and a 30-day fix order. You can run all of it yourself, for free.

Quick answer

To audit your AI search visibility, ask ChatGPT, Gemini, Perplexity, Claude and Google’s AI Mode 20 buyer questions (half with your brand name, half without), repeat each at least twice, and log whether you are named, described correctly and linked. Then score yourself on seven areas out of 100 using the rubric below. Fix the lowest-scoring plumbing first, and re-test every month.

What you are about to score, and why it works this way

We run this audit for clients before we touch anything else. It has six areas plus one baseline: AI citation, brand consolidation, structured proof, community and client proof, localisation, and pricing clarity, with on-site SEO as the reference point underneath. The score out of 100 is our own method. It is not an industry standard, and a high score does not guarantee a single recommendation. What it does is stop you guessing.

Four pieces of research shaped the weights. We want to be clear that none of them is a Voltik result.

Pew Research Center tracked 900 US adults in April 2025. When Google showed an AI summary, people clicked a normal result link in 8% of visits. Without one, they clicked in 15%. Some of your buyers now get an answer without ever visiting a site.
Ahrefs matched 15,000 long-tail prompts against Google’s results in August 2025. For ChatGPT, only about 8% of in-text citations pointed to a page in Google’s top 10 for the same query. Ranking well on Google does not carry over by itself.
Ahrefs also studied 75,000 brands in May 2025. How often a brand was mentioned around the web correlated with its appearance in Google AI Overviews at 0.664. Backlinks correlated at 0.218. That is correlation and not proof of cause. It still points off your own site.
SparkToro and Gumshoe.ai ran the same brand-recommendation prompts 60 to 100 times per platform in late 2025. The same list of brands came back less than 1% of the time. Top brands still appeared in 55% to 77% of answers in one category. So you measure how often you appear, never one answer.

The research behind the weights. These are studies by others, and most show correlation.

The 100-point scorecard. AI citation carries a quarter of the marks because it is the result everything else is meant to move.

Area	Points	What it asks
1. AI citation	25	Are you named when a stranger asks, and described correctly when they ask about you?
2. Brand consolidation	15	Does AI see one business, or several versions of it?
3. Structured proof	15	Can a machine verify who you are, who writes for you and what you have achieved?
4. Community and client proof	15	Do other people say you are good, in places AI reads?
5. Localisation	10	Does it know where you operate and recommend you there?
6. Pricing clarity	10	Is there a price signal a machine can quote?
7. On-site baseline	10	Can the bots get in and read the page?

Step 0: How do you set up an AI search test?

Use a clean setup so your own history does not flatter you. Open a logged-out browser window or a temporary chat in each tool, because saved memory and past chats can tilt answers toward businesses you have already discussed. Create a spreadsheet with the log columns below before you ask anything.

Tools to test: ChatGPT (with search on), Gemini, Perplexity, Claude (with web search on) and Google’s AI Mode or AI Overviews. If you sell in more than one country, add a VPN and repeat the unbranded prompts from each country. That is how we test localisation.

How much is enough? We run hundreds of prompts, competitors and locations for a client audit. You do not need that. Twenty prompts, three tools and two runs each is 120 answers, which is roughly two hours of copying and pasting. That is enough to see a pattern.

Common mistake: asking one question once and drawing a conclusion. The SparkToro result above is why. A single answer is a coin toss. A hundred answers is a measurement.

How do you run the prompt test?

Write 20 prompts in two halves, run each one twice per tool, and record every answer in the log. Ten prompts have no brand name in them (a stranger’s question). Ten include your name (a buyer checking you out before they commit).

The prompt test in six steps. The whole thing fits in one afternoon.

The prompt pack (copy and edit the brackets)

Without your brand name (these decide your Area 1 score):

What are the best [what you sell] in [city or country]?
Who should I use for [service] if I am a [type of buyer]?
Recommend a [supplier or provider] for [specific problem] in the UK.
Which [category] companies are best for [use case] and why?
What should I look for when choosing a [category] provider?
How much does [service] cost in the UK?
[Your biggest competitor] alternatives for [use case].
Best [category] for small businesses / for large companies / for [your niche].
Is it worth paying for [service], and who is good?
I need [outcome] by [deadline]. Who can help?

With your brand name (these decide your accuracy marks):

What is [Brand] and what do they do?
Is [Brand] any good? What do customers say?
How much does [Brand] cost?
[Brand] vs [competitor]: which is better for [use case]?
Where is [Brand] based and who runs it?
Does [Brand] work with [type of customer]?
What are the downsides of [Brand]?
Who founded [Brand] and what are their credentials?
Is [Brand] legitimate? Any complaints?
What is [Brand]’s best product or service?

Common mistake: writing prompts the way you describe your business. Write them the way a customer who does not know you would ask. Pull the wording from your enquiry emails, sales call recordings and the “People also ask” boxes on Google.

The log template

Make one row per answer. This is the whole data model, and you can paste it into a sheet today.

Date	Tool	Country	Prompt #	Run	Branded?	Named? (Y/N)	Position (1st, 2nd, listed)	Described correctly?	Linked to your site?	Who was named instead	Sources the AI cited

Two columns do most of the work. “Who was named instead” shows your real competitors, which are often not the ones you assumed. “Sources the AI cited” tells you which sites carry weight in your category, and that list becomes your off-site target list in Area 4.

The log template. Copy the columns into a spreadsheet before you start.

Area 1: AI citation (25 points)

What it measures: how often you are recommended when someone asks without your name, and whether AI gets the facts right when they ask with it.

How to score it. Twenty points come from the unbranded prompts. Appearance rate is the number of answers that name you divided by the number of unbranded answers.

Appearance rate	Points
0%	0
1% to 10%	5
11% to 25%	10
26% to 50%	14
51% to 75%	17
Over 75%	20

The last five points come from the branded prompts. Give yourself 1 point for each of these that is true across most answers: the description is correct, the location is correct, the services listed are correct, pricing or “from” prices are correct or honestly absent, and the answer links to your own site.

Worked example (illustrative, not a client). A UK supplier runs 10 unbranded prompts across 3 tools twice, so 60 answers. It is named in 9 of them. That is 15%, so 10 points. On the branded prompts the description and location are right, but the AI names a service they stopped selling in 2022 and links to a directory instead of the site. They score 2 of 5. Area 1: 12 out of 25.

What to fix, in order:

Read the “sources cited” column. If the AI keeps citing the same five sites for your category, those five sites are where your proof needs to be. Check which of them you appear on.
Where the AI states a wrong fact, find where it learned it. Search the wrong fact in quotes. It is usually an old directory entry, an old press page or an old version of your own site.
Publish one clear page per question buyers ask, with a direct answer in the first two sentences. Put the question as the heading, the answer straight underneath, and your evidence after that. That shape is easy for an AI tool to quote.

Common mistake: treating a high ranking position inside an answer as the goal. SparkToro’s finding is that appearing often is more stable than appearing first. Track appearance rate.

Area 2: Brand consolidation (15 points)

What it measures: whether AI sees one business or several. When facts are scattered, models can treat an old name, a second domain or a duplicate profile as a different company. In audits we call it a split source of truth, and we have found sites that AI mistook for a competitor because an old rebranded site was still live.

Give yourself 3 points for each of these five checks.

One domain. You run one main site, or if you have two markets, each is claimed, cross-linked and says it belongs to the same company. A second site that repeats your content splits your authority.
Old names are redirected. Search your previous brand name. Every old domain should return a 301 redirect to the new site, not a live copy.
Same name, address and phone everywhere. Compare your site, Google Business Profile, LinkedIn, Companies House, Trustpilot or other review sites, and your top five trade directories. Write the legal name and trading name the same way each time.
One description. Write a single 25-word description of what you do and who for. Use it in your site footer, LinkedIn, Google Business Profile and every directory. Models learn by repetition.
The profiles point home. Each profile links to your main site, and your site links back to each profile, so the connection works in both directions.

What to fix: build a one-page fact sheet (legal name, trading name, founded date, address, phone, one-line description, services, founders and roles, registration numbers). Paste from it every time, so nothing drifts. Then fix the loudest profiles first.

Common mistake: launching a second site on a new domain for a new service or campaign. Unless there is a strong reason, put it on a folder of the main site.

Area 3: Structured proof (15 points)

What it measures: whether a machine can verify that you are who you say you are. In one audit we found a firm’s best expert articles credited to a generic “advisors” byline. We called it asking Google to trust a ghost. A named, credentialed person is a fact a model can check. A brand with no one behind it is a claim.

Give yourself 3 points for each:

Named experts. Every important article has a real author with a bio page, credentials, registration number if you have one, and links to their LinkedIn profile.
Credentials in text. Accreditations, memberships, years trading and certifications are written as text on the page, not only inside a logo image.
Original evidence. You publish something only you could: customer numbers, test results, a survey, a teardown, a dated case study with before and after figures.
Schema that matches the page. You use Organization, Person, Article and FAQ markup where it fits. Google’s guidance is that structured data should match the visible text, so never mark up something a visitor cannot see. Microsoft’s Bing product team has said schema helps its LLMs understand content, as reported at SMX Munich in March 2025. Check yours with Google’s Rich Results Test and the Schema Markup Validator.
Dates. Pages show a published date and an honest updated date. Ahrefs looked at 17 million citations and found that URLs cited by AI assistants were on average 25.7% newer than those in organic results (1,064 days against 1,432), with ChatGPT showing the strongest preference for fresh content. Ahrefs’ own advice is that most brands gain more from new, good content than from endless refreshes.

What to fix: start with the three pages that matter most. For each, add a named author, a two-sentence credential line, one original data point and a visible updated date. Then repeat for the next ten.

Common mistake: adding schema to pages that have nothing worth marking up. Markup tells a machine what the page already says. It cannot invent proof.

Area 4: Community and client proof (15 points)

What it measures: whether other people say you are good, in places AI reads. This is the area that does the most for visibility and gets the least attention, because you cannot control it directly. Our view, backed by the Ahrefs mentions study above: those who skip off-site work will lose to those who do it.

Peec AI analysed about 30 million sources cited by AI tools in the US in March 2026. Reddit, YouTube, LinkedIn, Wikipedia, Forbes, G2, Yelp, Facebook, Medium and TechRadar led the list. Peec sells AI monitoring tools and says every industry has its own dominant sources, so use your own “sources cited” column as the real guide.

Score yourself:

Check	Points
Review volume and recency match your real customer volume. A handful of Google reviews, against thousands of customers served, looks thin to a machine and to a human.	4
You appear in trade press, industry bodies and the directories your category uses, and the entries are accurate and recent.	4
You or your team take part in communities where buyers ask (Reddit, forums, LinkedIn comments, Q&A sites) as real people, not as a link drop.	3
Named case studies or testimonials, with permission, say what you did and what changed.	4

What to fix:

Reviews. Set up a simple email or text sent a few days after delivery, asking for a review on the platform your category uses. A basic review-request automation is the cheapest fix, because happy customers rarely think to write one. Ask everyone, not only the happy ones, and never offer payment or incentives that platforms ban.
Mentions. Pick the five sites from your sources-cited column and work out how to earn a place on each: a contributed article, an expert quote, an entry in a roundup, a directory listing, a talk. People ask for these far less often than you would think.
LinkedIn. AI tools cite LinkedIn heavily. A founder profile that explains the business in plain words, plus regular posts, gives them something to find. Our LinkedIn guidance covers the detail.

Common mistake: buying a pile of low-quality directory links. Models read context, and a hundred empty listings add noise. Ten accurate, relevant ones help.

Area 5: Localisation (10 points)

What it measures: whether AI knows where you operate and recommends you there. A company can be strong in one country and invisible in the next, and you only see it by testing from there. This is why we use a VPN in client audits.

Give yourself 2 points for each:

You name every place you serve in text on the site, with a page for each major location or country.
Your address and service area are written the same way on the site, Google Business Profile and directories.
Prices, currency, delivery, tax and regulation details change by market where they should.
If you run more than one language or country version, you use hreflang tags so the right page reaches the right audience.
When you run your unbranded prompts through a VPN in each target country, you appear in at least one in four answers.

Worked example (illustrative, not a client). A manufacturer sells to UK and Polish buyers from one English-only page. Tested from London it appears sometimes. Tested through a Warsaw VPN with Polish prompts, it never does. Fix: a Polish page covering the same facts, a local address mention, and a Polish trade directory listing.

Common mistake: assuming one global page covers everything. Pages written for an audience get recommended to that audience.

Area 6: Pricing clarity (10 points)

What it measures: whether there is a price signal AI can quote. Buyers ask “how much does X cost?” more than most owners expect. If your site says “contact us” and a competitor says “from £X”, the competitor gets quoted and you get skipped. In our experience a missing price pushes the model to guess from third-party pages that are often out of date.

You do not have to publish a rate card. Publish what you can.

A “from” price or a typical range in text (4 points). “Most projects cost between £X and £Y” beats silence.
What the price includes (2 points). Packages or tiers with a short list of what is inside.
The factors that move the price (2 points). Size, volume, speed, complexity. This builds trust and gives the AI something quotable.
Consistency (2 points). The same range on your site, your directory listings and your Google profile.

What to fix: write a “How much does [service] cost?” page. Open with the range in the first sentence, then explain what changes it, then link to a quote form.

Common mistake: hiding prices to protect negotiating room. You can keep the room. A range still leaves you plenty of it, and it keeps you in the answer.

Area 7: On-site baseline (10 points)

What it measures: whether the bots can reach and read your pages at all. If this fails, nothing else on this list matters. Google’s own guidance for its AI features is plain: there are no additional requirements to appear in AI Overviews or AI Mode, and you do not need new machine-readable files, AI text files or special markup. The page must be indexed and eligible to show with a snippet. Basics are the baseline.

Check 1: Can the bots read your text? (3 points)

Several AI crawlers fetch your page without running the scripts that build it. Vercel and MERJ analysed this in late 2024 and reported that none of the major AI crawlers, including OpenAI’s and Anthropic’s, render JavaScript. If your content only appears after scripts run, they may see an empty shell. Test it in 60 seconds:

Open your key page in Chrome, right-click and choose View page source (not Inspect).
Press Ctrl+F (Cmd+F on Mac) and search for a sentence from your page.
If you cannot find it, the bots may not see it either. Give yourself 0. If it is there, 3.

In audits we see this most with page builders and template sites: legacy page builders on a CMS, template web builders, and SEO plugins that were never set up for AI crawlers. A pretty page is not a readable one, because machines see the code, not the design.

Check 2: Does robots.txt let the search bots in? (2 points)

Open yoursite.com/robots.txt. These are the crawlers that matter, taken from each company’s own documentation:

Crawler	Company	What it does	Our advice
Googlebot	Google	Crawls for Search, which feeds AI Overviews and AI Mode	Allow
OAI-SearchBot	OpenAI	Surfaces sites in ChatGPT search. Opt out and you will not appear in ChatGPT search answers	Allow
GPTBot	OpenAI	Collects content that may train OpenAI models. Separate from search	Your choice
ChatGPT-User	OpenAI	Visits a page when a user asks. OpenAI says robots.txt rules may not apply	Leave alone
Claude-SearchBot	Anthropic	Indexes content to improve Claude search results	Allow
Claude-User	Anthropic	Fetches pages when a Claude user’s question needs it	Allow
ClaudeBot	Anthropic	Collects public content that may be used for training	Your choice
PerplexityBot	Perplexity	Surfaces and links sites in Perplexity results. Not used for training	Allow
Perplexity-User	Perplexity	Fetches pages for a user’s question. Generally ignores robots.txt	Leave alone
Google-Extended	Google	Controls use in some Google AI training and grounding. Google says it does not affect inclusion in Search	Your choice

The useful point is that search access and training access are separate switches. You can allow the bots that put you in answers and block the ones that train models, if you want to. OpenAI says its search-related robots.txt changes can take about 24 hours to take effect.

A safe starting file:

User-agent: *
Allow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

Sitemap: https://yoursite.com/sitemap.xml

Also check the layer above robots.txt. Cloudflare announced in July 2025 that it was changing its default to block AI crawlers unless they pay, and sites on a CDN, security plugin or host firewall can block bots without any line in robots.txt. If you use one, look in its bot settings.

Check 3: Is the page clean? (5 points, 1 each)

Indexed. Google Search Console, URL Inspection, shows “URL is on Google” for your key pages.
One URL per page. Same content at /page and /page/ is two URLs. Pick one and add a canonical tag. We find this often, and it looks like duplication to a crawler.
No placeholder code. We have seen alt text about bed linen on a site that sold nothing of the kind, empty “Contact” links, and WhatsApp buttons with no number. This is site hygiene, and it affects trust and rankings.
Unique titles and one H1 per page. Titles under about 60 characters, descriptions under about 155.
A sitemap and fast, mobile-friendly pages. Submit the sitemap in Search Console. Consider adding your site to Bing Webmaster Tools as well, because Bing’s index feeds Copilot and other products.

The on-site baseline. If these fail, the other six areas have nothing to stand on.

A note on llms.txt. People ask about it constantly. The evidence is thin: see Does llms.txt work?. It earns no points in this scorecard.

How do you read your total score?

Add the seven areas. The bands below are ours, based on what we see when clients start.

Score	What it usually means	Where to start
0 to 39	Invisible. Often a readability or blocking problem, or a split brand	Area 7, then Area 2
40 to 64	Findable, rarely chosen. The facts are right, the proof is thin	Areas 3 and 4
65 to 84	Competitive. You show up, and competitors show up more often	Areas 4, 5 and 6, plus content
85 to 100	Strong. Defend it and keep the log running	Monthly re-test

The score is our own scale, and a target is a plan, not a promise. We cannot promise rankings, citations or leads, and nobody honest can. We can promise to measure the same way each month, so you see whether the work moves the number.

What is the 30-day fix order?

Fix the plumbing first, then the proof. We use the same order in our 90-day plans, where the first 30 days are the critical, high-return fixes.

The 30-day order. Each week builds on the one before, so resist skipping ahead.

Week 1: let the bots in and make the site readable.

Run the three on-site checks. Fix robots.txt, any CDN block and any page that fails the view-source test.
Fix duplicate URLs and placeholder code.
Submit the sitemap in Google Search Console and Bing Webmaster Tools.

Week 2: merge the brand.

Write the fact sheet and the 25-word description.
Redirect old domains, fix profiles and directories, and align name, address and phone.
Claim and complete Google Business Profile and LinkedIn.

Week 3: put names and proof on the page.

Add named authors, bios and credentials to your top 10 pages.
Publish the pricing page and one question-and-answer page for each of your top five buyer questions.
Add Organization, Person, Article and FAQ markup that matches the visible text, and validate it.

Week 4: start earning mentions and set up tracking.

Switch on the review-request message.
Pitch the five sites from your sources-cited column.
Re-run the prompt test and log it as your “Day 30” baseline.

Common mistake: starting with content because it feels productive. A new page on a site the bots cannot read adds nothing.

How do you keep tracking it?

Re-run the same prompts on the same tools on the same day every month, and keep the log. Compare three numbers: unbranded appearance rate, branded accuracy and your total score. Weekly checks add noise because the answers vary. Three months of data tell you far more than three weeks.

For a view over time, use a third-party tracker rather than building your own measurement. At Voltik we use external tools so we are not marking our own homework.

Be realistic about timing. Some fixes, like unblocking a crawler, can take effect within days. Reputation, reviews and mentions build over months. In our case study with MVPs, a sports memorabilia retailer, the starting point was a 0% AI citation rate and a #23 AI search rank, and the programme reached #2 on three of four platforms. That came from sustained work on content, off-site mentions and trust signals month after month, not from one fix. Another client, Uniflex, shows the website and content side of the same approach for a manufacturer.

Questions people ask

How do I check if ChatGPT recommends my business?

Open a logged-out window, turn search on, and ask the questions a customer would type, without your brand name. Do it for at least 10 questions and repeat each twice. Count how many answers name you. One answer tells you almost nothing because ChatGPT varies its replies.

How do I get my business recommended by ChatGPT?

Make sure OAI-SearchBot can crawl your site, make your text readable without JavaScript, keep your facts the same everywhere, and earn real mentions on the sites ChatGPT already cites in your category. Nobody can guarantee a recommendation. You can raise the odds and measure the change.

How long does AI search optimisation take?

Technical fixes can show up within weeks. Building reviews, mentions and authority usually takes months. We plan in 90-day blocks, with the first 30 days on high-return fixes, and report monthly.

Do I need schema markup for AI search?

Google says it does not require special markup for its AI features. Markup still helps machines understand a page, and Microsoft’s Bing team has said it helps its LLMs. Add it where it matches visible content, and treat it as support, not a shortcut.

Is AI search optimisation different from SEO?

They overlap on crawling, content and trust. They differ on where proof comes from and how you measure. See SEO vs AEO vs GEO for a full comparison.

Can I do this myself?

Yes. Everything above is designed so you can. The testing is easy. The slow parts are fixing the site and earning mentions elsewhere.

Sources

Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025.
Ahrefs, “Only 12% of AI Cited URLs Rank in Google’s Top 10 for the Original Prompt”, 11 August 2025.
Ahrefs, “An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied)”, 26 May 2025.
Ahrefs, “New Study: AI Assistants Prefer to Cite Fresher Content (17 Million Citations Analyzed)”, 28 July 2025.
SparkToro and Gumshoe.ai, AI recommendation consistency study, as reported by Search Engine Journal, 30 January 2026.
Peec AI, “Top domains cited by AI search: Analysis based on 30M sources”, March 2026.
Google Search Central, “AI features and your website”.
Google Search Central, “Google’s common crawlers” (Google-Extended).
OpenAI, “Overview of OpenAI Crawlers” (OAI-SearchBot, GPTBot, ChatGPT-User), developer documentation.
Anthropic, “Does Anthropic crawl data from the web, and how can site owners block the crawler?”, help centre.
Perplexity, “Perplexity crawlers” (PerplexityBot and Perplexity-User), documentation.
Vercel and MERJ, “The rise of the AI crawler”, 17 December 2024.
Cloudflare, “Content Independence Day: no AI crawl without compensation”, 1 July 2025.
Search Engine Land, “Microsoft Bing/Copilot use schema for its LLMs”, March 2025.

Want us to do this for you?

If you would rather have the audit run and the fixes made by people who do this every week, we do exactly that. Book a free call and we will show you your score, or read about our AI search optimisation service first.

AI Search Optimisation Checklist: The 6-Area Audit You Can Run Yourself (Score Out of 100)