HTML was conceived as a language for marking up a document that was primarily text; XML took the tags and attributes from that and tried to turn it into a data serialization and exchange format. But it was never really well suited to that, and it's obvious from looking XML-RPC or SOAP payloads that there were fundamental gaps in the ability of XML to encode type and structure information inline:

    <?xml version="1.0"?>
    <methodCall>
        <methodName>math.add</methodName>
        <params>
            <param>
                <value><int>5</int></value>
            </param>
            <param>
                <value><int>7</int></value>
            </param>
        </params>
    </methodCall>

Compared to this, JSON had string and number types built in:

    {
        "jsonrpc": "2.0",
        "method": "math.add",
        "params": [5, 7],
        "id": 1
    }

I don't think this is the only factor, but I think XML had a lot of this kind of cognitive overhead built in, and that gave it a lot of friction when stacked up against JSON and later yaml... and when it came to communicating with a SPA, it was hard to compete with JS being able to natively eval the payload responses.