Auto Alpha AdvisoryMy own audit engine scores autoalphaadvisory.co.za 88 out of 100. That's a keyed read from 27 September 2026. I ran it three times and got the same result each time, and the weakest of the nine domains is GEO at 70. Every automated check in the site's build passes.
On 10 October I had the same site read the way The Full Read reads a client's. I'm a CA(SA) who came up through audit, and the read works the way I was trained to: every claim against its source, every price against the price list, the terms against the marketing. That read produced 136 findings. Sixteen are rated High, and none of the 16 is something a scoring tool checks for.
If you found this by searching for a website audit in South Africa, here's the short answer to what The Full Read finds that a free scan doesn't. A scan tells you what machines can see on your site. The read tells you whether what the site says is true, and who four AI models name when they're asked your buyers' questions. The same limit applies to any review that scores a site against a checklist, free or paid. It finds what the checklist describes.
Sixteen findings a score had no way to see
A scoring engine holds your site against a rubric, with checks like whether the certificate is valid or whether the server sends a page's text without needing JavaScript. Those are good questions, and the scan answers them in minutes.
What it can't do is hold the site against anything outside it. It doesn't know which study a figure came from, what your current price list says, or what your terms commit you to. The read puts the site next to those documents and reads both. A score doesn't, and that's how a site can score 88 and still carry 16 High findings.
I'm not listing the 16 here, because I don't name a finding in a post until it's fixed on the live site. A post about checking your own claims that pointed readers at live errors would be a finding of its own.
Is the free scan still the right first step?
Yes, and it costs nothing. The free scan needs no call and no card. It scores nine domains, from security to AI readiness (GEO), and emails you a PDF report in minutes. It runs the checks that are tedious to do by hand, from security headers and page speed to whether your email domain can be spoofed. If those are broken, fix them before you pay anyone to read anything.
My scan page is plain about its limits: "A scan scores what a machine can see." It reads what the machines see. It doesn't read whether what you say is true.
That limit isn't peculiar to my engine. The W3C's Web Accessibility Initiative (WAI) says the same about accessibility tools in its guide to selecting evaluation tools, updated 13 May 2024: "Web accessibility evaluation tools can not determine accessibility, they can only assist in doing so." In the same section it says: "Human judgement is required." If a tool can't decide whether a page is accessible, it isn't going to decide whether a claim on that page matches its source.
The gap between a build that passes every gate and what a person sees on the page is its own story: 126 automated checks passed while every heading was invisible. This post is about whether what they read there is true.
What the read adds to the nine scores
The Full Read page lists what the read covers:
- the nine scored domains, the same ones the scan runs
- your public claims, each traced to its source: registers, filings and your own documents
- your terms and policies, read against your marketing
- who four AI models name when asked your buyers' questions, and who they name instead
- what each AI crawler is actually served, route by route, as each crawler identifies itself
- your regulatory position, read the way a regulator would
The two AI lines are the ones a scan comes closest to, so they're worth separating out.
Who four AI models name
The read puts your buyers' questions to four AI models, the API versions of the models behind ChatGPT, Claude, Gemini and Perplexity, and records which businesses come back. The scan checks whether your pages are readable by a machine. It doesn't put your buyers' questions to any model, so it can't tell you that a competitor gets named for your core service and you don't.
What each crawler is served
A robots.txt file tells you what you've allowed. It doesn't tell you what each crawler gets when it requests a page. Vercel's study with MERJ, published 17 December 2024, found that "none of the major AI crawlers currently render JavaScript." The crawlers it groups as ChatGPT and Claude fetched JavaScript files in 11.50% and 23.84% of their requests and didn't execute them, and the ChatGPT group spent 34.82% of its fetches on 404 pages, asking for addresses that weren't there. A page that builds its text in the browser can look finished to you and arrive without that text for them.
The crawlers don't all do the same job, either. OpenAI's crawler documentation describes GPTBot as collecting content that may be used to train its models, and OAI-SearchBot as surfacing sites in ChatGPT's search features. Anthropic's help page draws the same line between ClaudeBot (training) and Claude-SearchBot (search). Perplexity's bot documentation says PerplexityBot surfaces sites in its search results and isn't used to crawl content for foundation models. The read requests each route as each crawler identifies itself and records what each one is served.
Every finding comes with who can fix it
Each finding is classed by who can fix it: A for copy and config, B for engineering, C for anything that needs a decision or a third party, and D for what can't be fixed. The class tells you more than the count. On two full reads in September 2026, 35 of 58 and 39 of 56 findings were Class A or B, which the Full Read page describes as "the company's own team, hours to weeks."
So on those two reads, most of the list wasn't a consulting project. It was work the company's own people could schedule, plus a shorter set of decisions for whoever owns them. If you'd rather I did the fixing, the fix is proved by re-running the checks, finding by finding.
The Full Read is from R25,000, and the fee is credited in full against remediation that starts within 30 days. Because I came up through audit, I'm careful with the word: the read is not an audit or an assurance engagement, and it gives no assurance. What you get is the list, with every finding classed, before anyone else draws one up.
Someone else will read all of it in one sitting
Picture a buyer in due diligence, an investor with a term sheet, or the Information Regulator answering a complaint. Each of them reads the site, the terms, the claims and the filings together, in one sitting, looking for the place where two of them disagree.
That's the opportunity in a read done first, on your side. It finds those places while fixing them is still your decision, on your timetable. My site scores 88, and the read still turned up 16 High findings waiting for whoever read it closely. I'd rather be the one who found them.
If the scan's answers aren't the ones that worry you, the first call about a read is 15 minutes, and it starts on the Full Read page.