Writing for a web where the reader is often a machine

Where the site stands with AI assistants right now, what they actually come here for, how I would approach every new article from here, and what has been built underneath it all.

Now that the technical work is finished, this is the part that decides how far it carries. I pulled two weeks of raw server logs to see who is really reading the site, not estimates from a tool but every request the server answered. What came back changed what I would recommend, so I want to show you the evidence before the advice.

Where you standThe AI assistants are already here, in volume

This is every AI related crawler that touched the site over the past two weeks, alongside the traditional search engines for scale.

Who	What it does	Requests	Access
Bingbot	Bing search and Copilot	120,462	Allowed
Googlebot	Google Search and AI Overviews	61,559	Allowed
ChatGPT-User	ChatGPT fetching a page to answer someone	33,541	Allowed
Applebot	Siri and Spotlight	25,958	Allowed
OAI-SearchBot	ChatGPT's own search index	5,004	Allowed
DuckAssistBot	DuckDuckGo's AI answers	4,464	Allowed
Claude-SearchBot	Claude's search index	3,691	Allowed
Claude-User	Claude fetching a page to answer someone	214	Allowed
PerplexityBot	Perplexity's index	212	Allowed
ClaudeBot	Anthropic model training	580	Blocked
GPTBot	OpenAI model training	168	Blocked
Google-Extended	Gemini training	4	Blocked
CCBot, Bytespider, Amazonbot, Meta	Bulk scrapers and training	168	Blocked

The important thing here is the split, and it is working the way we want it to. Every crawler that can cite you is allowed in. Every crawler that only wants to train on you is turned away. ChatGPT, Claude, Perplexity, Copilot, DuckDuckGo and Google can all read and quote the site. What they cannot do is absorb it wholesale into a model that then never sends anyone back.

That distinction is the whole game, and it is worth being clear about because the two are easily confused. Blocking GPTBot does not remove you from ChatGPT. ChatGPT-User and OAI-SearchBot are the ones that matter there, and between them they made over 38,000 requests. The same is true for Claude: the training crawler is blocked, and Claude's search index still pulled 3,691 pages.

The demand signalWhat AI assistants actually come here for

This is the part I found genuinely useful. Because ChatGPT fetches a page live when a user asks something, the pages it requests are a direct readout of what people are asking it about your subject. Here are the most fetched:

CFA exam dates

141

CFA deferral and reschedule

137

CFA curriculum changes 2027

130

CFA exam cost

128

CFA refund and withdrawal policy

120

CFA curriculum changes 2026

109

CFA results date

103

Hedge fund career path

CPA exam dates and score release

CFA pass rates

There is a clear pattern in that list, and it is not the one I expected. Almost none of it is study advice or motivation. It is dated, factual, administrative information: when is the exam, what does it cost, can I defer, what changed this year, when do results land. The second cluster, further down, is career paths.

In other words, AI assistants are treating 300Hours as a reference source for facts that change every year. That is a strong position to hold, because those facts go stale annually and whoever maintains them most reliably becomes the default answer. It is fragile for exactly the same reason: the moment those pages drift out of date, the citation moves to whoever updated theirs.

The honest caveat

Being read is not the same as being visited. Over the same period, only around 50 people arrived here by clicking a link inside ChatGPT, Claude, Perplexity or Gemini, against roughly 12,000 from Google. The assistants are reading thousands of pages and answering in place.

I do not think that makes it worthless. Being named as the source in an answer builds the brand, and these tools increasingly show citations that people do click. But I would not plan on AI referral traffic replacing search traffic any time soon. Treat it as brand presence that is cheap to win because you are already most of the way there, not as a traffic channel to chase.

The current libraryWhat the existing 312 articles look like

I ran the whole published library through a check to see how well it already fits what these systems reward. Most of it is in good shape:

Freshness is excellent. 304 of 312 posts were updated within the last twelve months. For a site whose value is current facts, this is the single most important thing you already do well, and it should be defended above everything else here.
Depth is right. The median article is around 2,000 words. Only 19 sit under 800, long enough to be substantive without burying the answer.
Structure is mostly there. 305 of 312 use proper subheadings, and 42 per cent of those subheadings are phrased as questions, which is exactly the shape these systems match against.
Tables and lists are common. 61 per cent contain a table and all of them use lists. Both extract cleanly into an answer.

Two groups need attention, and I have listed every article in both below so they can be worked through directly.

Fix list one48 articles that bury the answer

These run more than 250 words before their first subheading. If someone asks what the CFA exam costs and the page opens with three paragraphs on the value of the charter, both the reader and the assistant have to dig for it. This is the highest value fix on the site and the cheapest.

Worth knowing before you start

30 of these 48 are CPA state requirement guides, which all share one template and one opening. That is not 30 problems, it is one template with 30 copies. Fixing the intro pattern once and applying it across the series clears nearly two thirds of this list in a single pass.

Read this before clicking that button

That button opens 48 separate browser tabs, one editor screen per article. It will slow the browser down noticeably and on some machines it will struggle. Your browser will almost certainly ask permission to open multiple tabs the first time, and you need to allow it or only the first tab appears.

Unless you are genuinely about to edit all of them in one sitting, use the individual Edit links in the table instead. You need to be logged in to WordPress for any of them to open properly.

Fix list two7 articles with no subheadings at all

These have no subheadings anywhere in the body. To a machine they are an undifferentiated wall of text, so nothing specific can be pulled from them, and to a reader on a phone they are equally hard going. All seven are provider reviews, so again there is a shared template underneath rather than seven unrelated problems.

A third item, worth noting but not worth a list: only 62 of 312 articles carry FAQ markup. That is the structured summary that lets a machine lift a clean question and answer pair without having to interpret your prose. On the fact heavy pages above it is close to free visibility.

Going forwardHow I would approach every new article

This is the part I would turn into a house rule. None of it is exotic and none of it means writing for machines at the expense of people. What makes an article easy for an assistant to quote is also what makes it easy for a stressed candidate to skim.

One article, one question

Resist the omnibus guide. A page that answers one question precisely gets cited. A page that answers nine gets skipped for all nine. If a piece needs to cover several things, they are several articles that link to each other.
Answer in the first forty words

Before context, before caveats, before why it matters. State the answer, then elaborate. The fix list above says this is where you are most often losing.

Not this: "Deciding to defer your CFA exam is never easy, and there are many factors..."
This: "You can defer your CFA exam once, to the next available window, for a $250 fee. Requests must be made before the registration deadline."
Date every fact that can expire

Write "as of August 2026" into the sentence itself, not just on the page. An assistant reading two conflicting sources favours the one that says when it was true. It also makes your annual refresh trivial, because you can search for the year.
Put the numbers in a table

Fees, dates, pass rates, weightings. Prose describing five numbers is far harder to quote accurately than five rows. You already do this on most pages, so make it the default.
Write subheadings as the question, in the reader's words

"How much does the CFA exam cost in 2026?" rather than "Costs". Use the phrasing candidates actually type, not the tidy editorial version.
Make every section stand alone

Assume each section will be read in isolation, because that is precisely what happens. The assistant takes the relevant chunk, not the article. Avoid "as mentioned above" and pronouns that reach back to an earlier section.
Say who says so, and link to them

Attribute to CFA Institute, GARP or AICPA by name and link the source page. It costs a click you might have kept and it buys the thing that makes an assistant willing to repeat you.
Add the FAQ block

Three to five real questions at the end of the fact heavy pieces. It takes minutes and it is the most directly machine readable thing on the page.
Publish something no one else has

This is the one that compounds. Your own candidate surveys, salary figures and study hour data are the reason an assistant names a source at all. Aggregated advice gets absorbed anonymously. Proprietary data gets attributed. If there is budget for one new content initiative a year, I would spend it here.

Where I would focusThe gaps worth filling

Keep the annual facts pages on a calendar. Exam dates, costs, deferral policy, curriculum changes, results dates, pass rates. These are your most cited pages by a wide margin. I would set a recurring reminder against each, timed to when the awarding body publishes, rather than updating reactively.
Extend the career path series. Hedge fund, wealth management and private equity all appear high in the fetch list, which suggests appetite well beyond what is published. That cluster is doing more work than its size suggests.
Add direct comparisons. CFA against CFP, FRM against CFA, CPA against CFA for a given background. Comparison questions are disproportionately common in assistants because they are hard to answer from any single official source, which is exactly the gap a good independent site fills.
Cover CPA and FRM logistics as thoroughly as CFA. CPA exam dates already ranks highly on far less material. The CFA pages are proven, and the same treatment on the other qualifications is a known formula on untapped ground.

Under the bonnetWhat has been built

Quite a lot has changed underneath the site over this engagement. This is what is now in place and what each piece actually does, so you know what you are working with.

The advertising system

Sponsor banners used to be written directly into the theme and into widgets, which meant every change to an advert was a developer job. They are now managed from a single Global Ads screen in the WordPress back end, and the team can run it without me.

Advertisers are set up as clients, and each client holds its own list of adverts, so everything for one sponsor sits together in one place.
Category targeting decides where an advert appears. This is now hierarchical, which matters more than it sounds: an advert aimed at CFA automatically shows on CFA Level 1, Level 2 and Level 3 articles as well, instead of silently missing them.
Country targeting uses the visitor's country as supplied by Cloudflare, so a sponsor who only sells in one region is not paying for impressions elsewhere.
Rotation shares a slot between several adverts rather than one banner holding a position permanently.
Impression counting records what was actually seen, broken down by advert, by day and by country. As of now that is 9,364 impressions across 26 adverts and 88 countries, readable from a reporting screen in the back end.
Every advert has a permanent internal identifier. This was a real risk that has been closed: the counting used to be keyed on the advert's name and its position on the page, so renaming a banner would have started its history from scratch, and handing a slot to a new sponsor would have quietly merged two advertisers' numbers into one figure that could never be separated again. Reporting now follows the advertiser rather than the slot. Your existing Google Analytics tracking is untouched, since the identifiers the analytics uses were deliberately left exactly as they were.

The video player

There is now a video that travels down the left hand column with the reader instead of scrolling away at the top of the page.

A site wide default video is set once in the back end, and any individual article can override it with its own.
As you scroll, the player docks to the bottom of the left column and stays there, remaining visible for the whole article. It measures the column on every scroll, so it stays properly aligned rather than drifting sideways, and it reserves its own space so it never sits on top of the navigation above it.
It carries an expand control and an optional call to action button underneath, which is set alongside the video itself.
The player is loaded lazily, meaning the video is only requested when it is actually needed. A video embed is one of the heaviest things you can put on a page, and this keeps it from slowing the article down.

The automatic left hand navigation

The left column builds itself. Nothing has to be maintained by hand, and it adapts to what kind of page the reader is on.

On an article or a hub page it reads the actual headings on the page and builds an "On this page" contents list from them, numbered and linked, so a reader can jump straight to the section they came for. Because it is generated from the live page, it can never fall out of sync with the article the way a hand written list does.
On a category page or an article feed, where there are no headings to work with, it switches to a "Most Recent" list of the articles in that section instead. The previous approach produced an empty column on these pages.
It tracks your position as you scroll and highlights the section you are currently reading.
The three column layout itself was extended to cover articles, hub pages, article feeds and category archives. The home page is deliberately excluded, since it is a landing page with its own composition and the reading layout is the wrong shape for it.
This replaced a plugin generated contents box that was duplicated on 288 articles, showing readers the same list twice.

Cloudflare

Cloudflare was originally put in front of the site for protection and speed, and that remains its main job. It sits between visitors and the server, so it absorbs spam traffic, bot attacks and abusive requests before they ever reach the site, and it serves your images and files from locations near the reader rather than from the single server. A forum the size of yours attracts a constant stream of automated registration and posting attempts, and this is what keeps that away from the server.

On top of that original setup, three things were added:

The AI crawler policy described at the top of this document, separating crawlers that cite you from crawlers that only harvest you.
An llms.txt file, which is the emerging convention for telling AI systems how a site is organised and what it is authoritative on.
The country detection that the advert targeting relies on, which Cloudflare provides at no extra cost.

Search engine groundwork

The site was moved from Rank Math to Yoast, with all titles, descriptions and settings carried across and checked against a baseline of live pages before and after.
FAQ markup was repaired across 81 blocks and 557 questions that had been silently broken and were producing no structured data at all. They looked fine on the page while being invisible to search engines.
Breadcrumbs were rebuilt and now show the category path on every article.
The HTML sitemap was cut from 2,627 links to 620, so it describes your articles instead of drowning them in forum threads.
Featured images were given correct sizing and loading priority, which is one of the things Google measures directly.

Still outstandingWhat should be cleaned up next

None of this is urgent and none of it is breaking anything today, but it is all worth doing and it is the kind of thing that only gets more awkward the longer it waits.

Item	Detail	Priority
Orphaned user accounts	There are 54,897 user accounts on the site, of which 44,519 have no role assigned at all and 51,897 have never posted anything. They are almost entirely automated registrations from before the sign up button was removed: the registration count runs at 15,808 in 2022 and 14,673 in 2024, then collapses to 301 in 2025 and 1 so far in 2026.	High
Plugin updates	14 plugins have updates waiting. The block editor plugin, GenerateBlocks, is on 2.1.0 against a current 2.4.0. Two others matter more: Gravity Forms is a full major version behind at 2.9.11 against 3.0.1, and Wordfence, which is your security plugin and the one that most needs to be current, is on 8.1.0 against 8.2.2.	High
Unused themes	Five WordPress default themes are installed and inactive (Twenty Twenty-One through Twenty Twenty-Five), all with updates pending. Inactive themes still need patching, so they are pure maintenance overhead. Standard practice is to keep one as a fallback and remove the rest.	Medium
Forum data	The forum still holds 3,696 topics and 21,105 replies, including 563 marked as spam, pending or trashed. If the forum is being retired, this needs a deliberate plan for redirects and archiving rather than a straight deletion.	Medium

On removing those 44,519 accounts

This one needs saying clearly, because the obvious approach will not work. Deleting them through the Users screen in WordPress is not enough and should not be attempted.

Every account carries a set of associated records behind it, and the site is currently holding 870,866 of them. Removing accounts through the admin screen is extremely slow at this volume, it will time out long before it finishes, and depending on how it fails it can leave the associated records behind entirely, so the database ends up no smaller while the accounts are gone.

It needs to be done directly against the database, in batches, with a full backup taken first and a check afterwards that no genuine account with real activity has been caught up in it. It is not a difficult job, but it is not a click in the admin either. Happy to handle it.

Living with itWhat to look out for

The site is stable. Everything described here has been tested on the live site and checked in a browser rather than just assumed to work. But a lot changed in a short space of time, and I would rather tell you where the edges are than have you discover them. If you see any of the following, it is very likely one of these causes rather than anything serious:

A change you made does not appear on the site. This is far and away the most likely thing you will hit, and it is almost never a real fault. The site runs aggressive caching, so visitors can be served a stored copy of a page for some time after it changes. Clear the cache before assuming something is broken. This caught me out more than once during the work.
A page loads with its styling missing or scrambled. The optimisation layer strips out stylesheet rules it believes are unused, and it occasionally gets that judgement wrong on a page type it has not seen before. Clearing the cache normally resolves it. Please tell me if a specific page does it repeatedly.
A new page type does not get the three column layout. The layout is applied deliberately to articles, hub pages, feeds and category archives, and deliberately not to the home page. If you add a new kind of page and it comes out in the old single column, that is the rule doing what it was told rather than a fault, and it is a small change to include it.
An advert appears somewhere unexpected. Targeting now follows the category tree, so an advert aimed at a parent category will correctly appear on its children too. If something looks out of place, check the article's categories first, since that is usually the explanation.
The video sits slightly off, or overlaps something. It positions itself by measuring the column live. Very unusual window sizes, or a browser zoom setting well away from 100 per cent, are the situations most likely to put it out.
The email sign up form looks wrong or appears in the wrong place. It only shows to logged out visitors, so you will not see it while signed in to WordPress. It should sit centred in the main column below the title. To check it, open an article in a private browsing window.
An FAQ section stops showing up in search results. The FAQ blocks store their questions and answers in a structured form behind the scenes, not merely as visible text. Editing one through raw HTML rather than the normal block editor can break that structure while leaving the page looking perfectly fine, which is precisely the fault that was found and repaired on 81 blocks.

None of these are likely, and I would expect a quiet few weeks. If anything does look wrong, the most useful thing you can send me is the page address and a screenshot, and I will take a look and get a fix applied.

The bigger pictureThe rebuild proposal

Everything above is about getting the most out of the site as it stands today. Separately, I put together a proposal covering the larger rebuild: the custom theme, reducing the plugin load, the forum question, the advertising system and the reporting dashboard, along with what each phase would involve and what it would cost.

Read the rebuild proposal. It is open for comments, so please add anything you want raised directly against the relevant section rather than collecting it up separately.

In shortThe checklist

If nothing else here survives, this is what I would pin above the desk for every new piece:

One question per article, asked in the reader's own words
The answer stated in the first forty words
Every expiring fact dated in the sentence itself
Numbers in a table, not in prose
Subheadings phrased as questions
Every section readable on its own
Sources named and linked
An FAQ block on anything fact heavy
A diary date for the annual refresh

Happy to go through any of this properly. I am glad to take the first pass at the 48 articles that bury the answer, since 30 of them share a template and it is largely mechanical, and to handle the account cleanup safely against the database.

Dino