Six sections: site health, tracking, messaging, Core Web Vitals, citation share and share of voice. The last two are the ones most audits still skip. Pew found people clicked a traditional search result in 8% of visits that showed a Google AI summary. A free audit sees the outside of all six, but not your conversion data.
Four of the six are old and two are new
The first four sections are familiar. Is the site technically sound? Does tracking record what actually happens? Does the page say what the buyer needs to hear, and does it load fast enough for them to read it? Any decent audit has covered these for years, and they still decide most of what your budget buys.
The two new sections exist because search changed shape. Citation share asks how often AI answers, in ChatGPT, Google's AI Overviews and the other assistants, name your brand or link to you when a buyer asks a question you should own. Share of voice asks the older question across every surface: when your category comes up, how much of the conversation is yours.
The case for skipping them is not stupid. Citation share is noisy, the tools are young, and nobody can yet tie an AI mention to revenue the way you can tie a paid click to a lead. I accept all of that. I still think an audit without them only measures the part of search you already understand. Pew Research Center found that Google users who saw an AI summary clicked a traditional result in 8% of visits. If some of your buyers read the answer before they see your link, the audit has to read the answer too.
Three numbers behind the checklist
Is the site readable by people and by crawlers?
Site health
Site health is the dull section and the one where an outside audit is strongest, because nearly all of it is public. Broken pages, redirect chains, duplicate titles, pages kept out of the index by accident, sitemaps that still list pages you deleted. A crawler finds these in an afternoon, and most of them are cheap to fix.
The 2026 addition is the crawler list. Your robots file now decides whether AI search tools can read you at all. OpenAI Developers puts it plainly: sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. It is easy to opt out by accident, through a blanket rule a developer added or a security plugin switched on, with nobody in marketing aware of it. An audit should list every AI crawler your site allows and blocks, and say whether each one was a decision or an accident.
Core Web Vitals
Google's guidance on web.dev classifies a 75th percentile LCP of 2 seconds as good. LCP is the time until the largest piece of content on the page appears. Two details in that threshold matter for an audit. It is measured on real visits, so a lab test from a fast office connection tells you little. And it is the 75th percentile, so the score is set by your slower visitors, usually on mobile, often in a market far from your servers.
Why speed earns its own section: the BBC found it lost an additional 10% of users for every extra second its site took to load. For a reference from our own work, the Berlitz Digital School site launched with a Lighthouse performance score of 94, averaging 2.1 seconds to load globally. Read that carefully, because an audit should too. Lighthouse is a lab score, and a global average load time is a different measure from LCP at the 75th percentile. A good lab number is where the check starts.
Does your tracking still count the conversions?
If I could pay for one section only, it would be this one. Every other number in the audit is read through your tracking. When tracking is wrong, the rest of the report is confident about the wrong thing.
The failures that show up most often are not exotic:
- Consent gaps. Where a consent banner sits in front of your tags, Google's consent mode can model some of the missing conversions. The condition is in the documentation from Google for Developers: your tag needs to meet a certain data collection threshold. Below it, you get no modelled numbers, and I suspect many smaller accounts sit below it without anyone noticing.
- Conversions counted twice or on the wrong event, such as a thank-you page that fires again on refresh, or a form start recorded as a lead.
- Three systems, three answers. The ad platform, analytics and the CRM each report a different lead count, and nobody owns the job of saying which is right.
An audit should end this section with one sentence: the number we trust for leads or revenue, where it comes from, and roughly how wrong it might be. If the auditor cannot write that sentence, that is the finding, and it outranks everything else in the report.
The fair objection is that tracking is never perfect, so spending on it has no end. I agree it is never perfect. The goal is a number whose error you know, which is cheaper to reach than a precise number and nearly as useful for deciding where money goes.
Messaging: does the page answer the buyer's real objection?
Messaging is where most audits go vague, because it is hard to score. So they score what is easy: headline length, a keyword in the main heading, a reading-age score, a button above the fold. None of that is wrong. It rarely explains why a page does or does not sell.
What a messaging audit checks, and what it should check
Headline length, a keyword in the main heading, a reading-age score, a button above the fold. Quick to tick off and rarely the reason a page fails.
Whether the page answers the objection that stops people buying, gives a reason to act now, and offers a route for people who will not fill in a form.
The best evidence I have for scoring substance over format comes from a Berlitz UK offer campaign, where 21% of leads became paid enrolments. The two decisions behind it were the closing date in the page bar and a WhatsApp route beside the form. Neither would have registered on a reading-age score. Both came from asking what would make a hesitant buyer act this week instead of next month.
A messaging section I would trust takes the pages with the most paid and organic traffic, writes down the one question each page must answer for its buyer, and says plainly whether it does. It adds one check that older audits never needed: does the AI description of your product match your own page? If an assistant describes your offer with old prices or a feature you dropped, the problem sits off your site, and only the messaging section will catch it.
How do you measure citation share and share of voice?
Start with citation share, because it is the less familiar of the two. The method is simple. The discipline is in repeating it the same way.
A citation share check you can repeat
Named and cited are different counts on purpose. A mention builds awareness. A citation with a link can send a visit. An audit that merges them flatters the result.
Here is a real baseline, with its limits stated. It was measured with CrawlerPal, a visibility tool from our own studio, so this is our measurement of our client. For Berlitz Canada we ran 28 prompts in ChatGPT, one run per prompt, before any new comparison content shipped. 26 of them left out the brand name. On those, Berlitz was named in 5 answers (19%) and cited in 4 (15%). That is a single run and the re-test is not published yet, so treat it as an example of the method rather than proof of anything. On the same account, the comparison articles written for AI answers also brought 115 Google clicks over three months, so writing for answers did not cost the ordinary search channel anything we could see.
Share of voice is the older measure and the easier one to fake, usually by counting impressions. Reach alone misleads. On the Berlitz Canada social programme, total impressions fell 38% in a year while the engagement rate per impression rose from 0.7% to 1.3%. An audit reading only reach would have called that year a decline. For an audit I would define share of voice as: in the searches, AI answers and feeds where your category is discussed, how often you appear compared with a short, named list of competitors. Naming the competitors in advance matters more than which tool counts them.
What a free audit can see, and what it cannot
We offer a free growth audit, so read this section as disclosure. Here is what any free audit, ours included, can and cannot do.
From the outside, an auditor can see almost all of site health, your crawler rules, public speed data where Google has enough visits to publish it, the messaging on every public page, and a fair first read of citation share and share of voice. That is a lot. It is usually enough to tell you where to look.
What it cannot see is everything behind your login. Whether conversions are counted correctly. How consent is configured beyond what loads in a browser. Which pages actually produce revenue. How the ad accounts are structured and what each click costs. Your Search Console data, which is the only reliable view of your own search clicks. So a free audit is good at pointing at leaks and poor at pricing them.
Four ways to buy an audit
The sensible sequence is the second cell. Take a free outside read first, from our free growth audit or from anyone else, and judge it on one thing: did it find something specific you did not know? If it did, give read-only access to analytics, tag manager, Search Console and a CRM export, and pay for the tracking section to be done properly. That is the part a free audit cannot reach, and the part most worth paying for.
Questions people ask
What is the difference between a free marketing audit and a paid one?
A free audit reads what is public: site health, crawler rules, speed data, messaging and a first pass at citation share and share of voice. A paid audit adds access to your analytics, tag manager, Search Console, ad accounts and CRM, which is where tracking errors and the cost of each leak become visible. The free one tells you where to look. The paid one tells you what the problem is costing.
How often should we audit our digital marketing?
A full audit once a year is reasonable for most companies. Re-check tracking after any change to the site, the consent banner or the CRM, because that is when counts break quietly. Citation share needs more frequent runs than the rest, since AI answers change between runs and a single sample proves little. Quarterly is a sensible starting rhythm, adjusted once you see how much your answers move.
Can I measure citation share in AI answers myself?
Yes. Write the questions your buyers actually ask, mostly without your brand name in them. Run each one in ChatGPT and any other assistant your buyers use, on the same day, and record the date. Count two things separately: answers that name you, and answers that link to you. Repeat on a schedule. The method is simple. The hard part is running it the same way every time.
What access should we give a marketing auditor?
Read-only access is enough for almost everything: analytics, tag manager, Search Console, the ad accounts, and an export of leads from the CRM for a recent period. Avoid giving edit rights for an audit. If an auditor asks for edit access before showing any findings, ask why. The tracking section cannot be done properly without this access, so a request for read-only access is a good sign.
Is a good Lighthouse score enough for Core Web Vitals?
No. Lighthouse runs a lab test on one device and one connection. Core Web Vitals are judged on real visits at the 75th percentile, so slow phones and distant markets count against you. A site can score well in the lab and still fail in the field. Use Lighthouse to find problems, and use real-visit data to decide whether you actually have one.
Where this is weak
Most of the evidence here comes from one client family, Berlitz, across a few markets. I think the checklist holds elsewhere, but I have not shown it. Citation share is the weakest section: a single run of prompts is a sample, answers change between runs, and nobody I know can yet link an AI mention to revenue with confidence. If your buyers do not ask assistants about your category, it may not earn its place yet.
Six sections can also be too many. A small business with one landing page and no paid spend should audit tracking and messaging and stop there. And I run a free audit that can lead to paid work, so my view on where the free part ends is not neutral.